{
  "id": 120015,
  "title": "Algorithm Selection for Beginner!",
  "url": "/competitions/pku-autonomous-driving/discussion/120015",
  "author_name": "",
  "post_date": "2019-12-03T05:24:58.397050500Z",
  "votes": 77,
  "comment_count": 39,
  "views": 0,
  "content": "<h3>Upvote if you like, thanks.</h3>\n\n<p>This is a big picture of <strong>object detection</strong>.</p>\n\n<p>The algorithm of object detection is devided into two types - <strong>1-stage</strong> and <strong>2-stage</strong>.</p>\n\n<h2>1-stage</h2>\n\n<ul>\n<li>anchor_based\n<ul><li><a href=\"https://pjreddie.com/media/files/papers/YOLOv3.pdf\">YOLO_v3</a></li>\n<li><a href=\"http://arxiv.org/abs/1512.02325\">SSD</a>-<a href=\"https://github.com/amdegroot/ssd.pytorch\">github</a></li></ul></li>\n<li>anchor-free\n<ul><li><a href=\"https://arxiv.org/abs/1808.01244\">CornerNet</a></li>\n<li><a href=\"https://arxiv.org/pdf/1904.07850.pdf\">CenterNet</a>-<a href=\"https://github.com/xingyizhou/CenterNet\">github</a></li>\n<li><a href=\"https://arxiv.org/pdf/1904.01355.pdf\">FCOS</a></li></ul></li>\n</ul>\n\n<h2>2-stage</h2>\n\n<ul>\n<li>RPN (Region Proposal Network) + anchor\n<ul><li><a href=\"https://arxiv.org/abs/1506.01497\">Faster R-CNN</a>-<a href=\"https://github.com/endernewton/tf-faster-rcnn\">github</a></li>\n<li><a href=\"https://arxiv.org/abs/1703.06870\">Mask R-CNN</a>-<a href=\"https://github.com/facebookresearch/Detectron\">github</a></li></ul></li>\n</ul>\n\n<p>I just list some basic algorithm, you can google it to learn more.</p>\n\n<h3>Finally, I'll give you some references which are basically same task with this competition.</h3>\n\n<ul>\n<li><a href=\"https://arxiv.org/pdf/1711.10006.pdf\">SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again</a>\n[Github link] (<a href=\"https://github.com/wadimkehl/ssd-6d\">https://github.com/wadimkehl/ssd-6d</a>)</li>\n<li><a href=\"https://arxiv.org/abs/1902.04103\">Bag of Freebies for Training Object Detection Neural Networks</a></li>\n<li><a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120510\">DeepIM: Deep Iterative Matching for 6D Pose Estimation</a></li>\n</ul>\n\n<p>This post will be updated as I find new useful references.</p>\n\n<h3>This is my other topics, hope you like.</h3>\n\n<p><a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/123385\">The coordinate system of this competition</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120076\">The algorithm that baidu apollo chooses!</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120083#latest-686785\">Understant Camera Intrinsic Parameters and X, Y, Z</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120381#latest-688477\">Algorithm Selection in 6D Pose Estimation</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120443\">How the Competition Organizer get train.csv !</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120653\">A way to improve the accuracy of model</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120710\">Object Detection in 20 Years</a></p>\n\n<p>My notebook:\n<a href=\"https://www.kaggle.com/diegojohnson/centernet-objects-as-points\">Centernet - Obejects as Points</a>\n<a href=\"https://www.kaggle.com/diegojohnson/a-way-to-regress-translation-and-rotation\">A Way to Regress Translation and Rotation</a>\n<a href=\"https://www.kaggle.com/diegojohnson/best-algorithm-so-far-google-s-new-paper-in-nov\">EfficientDet-D1</a>\n<a href=\"https://www.kaggle.com/diegojohnson/a-better-way-to-regress-yaw\">Regress angle in 2 bins - centernet way</a>\n<a href=\"https://www.kaggle.com/diegojohnson/a-clear-view-of-car-pose\">A clear view of car pose</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/124480\">Dataset with preprocess</a></p>",
  "messages": [
    {
      "id": "686395",
      "postDate": "12/03/2019 05:24:58",
      "content": "<h3>Upvote if you like, thanks.</h3>\n\n<p>This is a big picture of <strong>object detection</strong>.</p>\n\n<p>The algorithm of object detection is devided into two types - <strong>1-stage</strong> and <strong>2-stage</strong>.</p>\n\n<h2>1-stage</h2>\n\n<ul>\n<li>anchor_based\n<ul><li><a href=\"https://pjreddie.com/media/files/papers/YOLOv3.pdf\">YOLO_v3</a></li>\n<li><a href=\"http://arxiv.org/abs/1512.02325\">SSD</a>-<a href=\"https://github.com/amdegroot/ssd.pytorch\">github</a></li></ul></li>\n<li>anchor-free\n<ul><li><a href=\"https://arxiv.org/abs/1808.01244\">CornerNet</a></li>\n<li><a href=\"https://arxiv.org/pdf/1904.07850.pdf\">CenterNet</a>-<a href=\"https://github.com/xingyizhou/CenterNet\">github</a></li>\n<li><a href=\"https://arxiv.org/pdf/1904.01355.pdf\">FCOS</a></li></ul></li>\n</ul>\n\n<h2>2-stage</h2>\n\n<ul>\n<li>RPN (Region Proposal Network) + anchor\n<ul><li><a href=\"https://arxiv.org/abs/1506.01497\">Faster R-CNN</a>-<a href=\"https://github.com/endernewton/tf-faster-rcnn\">github</a></li>\n<li><a href=\"https://arxiv.org/abs/1703.06870\">Mask R-CNN</a>-<a href=\"https://github.com/facebookresearch/Detectron\">github</a></li></ul></li>\n</ul>\n\n<p>I just list some basic algorithm, you can google it to learn more.</p>\n\n<h3>Finally, I'll give you some references which are basically same task with this competition.</h3>\n\n<ul>\n<li><a href=\"https://arxiv.org/pdf/1711.10006.pdf\">SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again</a>\n[Github link] (<a href=\"https://github.com/wadimkehl/ssd-6d\">https://github.com/wadimkehl/ssd-6d</a>)</li>\n<li><a href=\"https://arxiv.org/abs/1902.04103\">Bag of Freebies for Training Object Detection Neural Networks</a></li>\n<li><a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120510\">DeepIM: Deep Iterative Matching for 6D Pose Estimation</a></li>\n</ul>\n\n<p>This post will be updated as I find new useful references.</p>\n\n<h3>This is my other topics, hope you like.</h3>\n\n<p><a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/123385\">The coordinate system of this competition</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120076\">The algorithm that baidu apollo chooses!</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120083#latest-686785\">Understant Camera Intrinsic Parameters and X, Y, Z</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120381#latest-688477\">Algorithm Selection in 6D Pose Estimation</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120443\">How the Competition Organizer get train.csv !</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120653\">A way to improve the accuracy of model</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120710\">Object Detection in 20 Years</a></p>\n\n<p>My notebook:\n<a href=\"https://www.kaggle.com/diegojohnson/centernet-objects-as-points\">Centernet - Obejects as Points</a>\n<a href=\"https://www.kaggle.com/diegojohnson/a-way-to-regress-translation-and-rotation\">A Way to Regress Translation and Rotation</a>\n<a href=\"https://www.kaggle.com/diegojohnson/best-algorithm-so-far-google-s-new-paper-in-nov\">EfficientDet-D1</a>\n<a href=\"https://www.kaggle.com/diegojohnson/a-better-way-to-regress-yaw\">Regress angle in 2 bins - centernet way</a>\n<a href=\"https://www.kaggle.com/diegojohnson/a-clear-view-of-car-pose\">A clear view of car pose</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/124480\">Dataset with preprocess</a></p>",
      "rawMarkdown": "### Upvote if you like, thanks. \n\nThis is a big picture of **object detection**.\n\nThe algorithm of object detection is devided into two types - **1-stage** and **2-stage**.\n\n## 1-stage\n- anchor_based\n    * [YOLO_v3](https://pjreddie.com/media/files/papers/YOLOv3.pdf)\n    * [SSD](http://arxiv.org/abs/1512.02325)-[github](https://github.com/amdegroot/ssd.pytorch)\n- anchor-free\n    * [CornerNet](https://arxiv.org/abs/1808.01244)\n    * [CenterNet](https://arxiv.org/pdf/1904.07850.pdf)-[github](https://github.com/xingyizhou/CenterNet)\n    * [FCOS](https://arxiv.org/pdf/1904.01355.pdf)\n\n## 2-stage\n-  RPN (Region Proposal Network) + anchor\n    * [Faster R-CNN](https://arxiv.org/abs/1506.01497)-[github](https://github.com/endernewton/tf-faster-rcnn)\n    * [Mask R-CNN](https://arxiv.org/abs/1703.06870)-[github](https://github.com/facebookresearch/Detectron)\n\nI just list some basic algorithm, you can google it to learn more.\n\n### Finally, I'll give you some references which are basically same task with this competition.\n\n* [SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again](https://arxiv.org/pdf/1711.10006.pdf)\n[Github link] (https://github.com/wadimkehl/ssd-6d)\n* [Bag of Freebies for Training Object Detection Neural Networks](https://arxiv.org/abs/1902.04103)\n* [DeepIM: Deep Iterative Matching for 6D Pose Estimation](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120510)\n\nThis post will be updated as I find new useful references.\n\n### This is my other topics, hope you like.\n[The coordinate system of this competition](https://www.kaggle.com/c/pku-autonomous-driving/discussion/123385)\n[The algorithm that baidu apollo chooses!](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120076)\n[Understant Camera Intrinsic Parameters and X, Y, Z](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120083#latest-686785)\n[Algorithm Selection in 6D Pose Estimation](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120381#latest-688477)\n[How the Competition Organizer get train.csv !](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120443)\n[A way to improve the accuracy of model](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120653)\n[Object Detection in 20 Years](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120710)\n\n\nMy notebook:\n[Centernet - Obejects as Points](https://www.kaggle.com/diegojohnson/centernet-objects-as-points)\n[A Way to Regress Translation and Rotation](https://www.kaggle.com/diegojohnson/a-way-to-regress-translation-and-rotation)\n[EfficientDet-D1](https://www.kaggle.com/diegojohnson/best-algorithm-so-far-google-s-new-paper-in-nov)\n[Regress angle in 2 bins - centernet way](https://www.kaggle.com/diegojohnson/a-better-way-to-regress-yaw)\n[A clear view of car pose](https://www.kaggle.com/diegojohnson/a-clear-view-of-car-pose)\n[Dataset with preprocess](https://www.kaggle.com/c/pku-autonomous-driving/discussion/124480)",
      "votes": null
    },
    {
      "id": "686399",
      "postDate": "12/03/2019 05:28:44",
      "content": "<p>If you want more references, see this post. <a href=\"/bibek777\">@bibek777</a> \n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/113895#latest-685345\">https://www.kaggle.com/c/pku-autonomous-driving/discussion/113895#latest-685345</a></p>",
      "rawMarkdown": "If you want more references, see this post. @bibek777 \nhttps://www.kaggle.com/c/pku-autonomous-driving/discussion/113895#latest-685345",
      "votes": null
    },
    {
      "id": "686482",
      "postDate": "12/03/2019 07:53:08",
      "content": "<p>following</p>",
      "rawMarkdown": "following",
      "votes": null
    },
    {
      "id": "686515",
      "postDate": "12/03/2019 08:39:30",
      "content": "<p>Thanks</p>",
      "rawMarkdown": "Thanks",
      "votes": null
    },
    {
      "id": "687170",
      "postDate": "12/04/2019 02:17:47",
      "content": "<p>Thanks for sharing. Very useful to     me.❗️</p>",
      "rawMarkdown": "Thanks for sharing. Very useful to     me.❗️",
      "votes": null
    },
    {
      "id": "687175",
      "postDate": "12/04/2019 02:28:29",
      "content": "<p>Wow, you have been 43rd, which algorithm do you choose?🤔 </p>",
      "rawMarkdown": "Wow, you have been 43rd, which algorithm do you choose?🤔",
      "votes": null
    },
    {
      "id": "687306",
      "postDate": "12/04/2019 07:57:36",
      "content": "<p>Hi\nI tried to use YOLO on custom data-set but it took very long time on few images. Is it take that much time or there may be some problem from myside?</p>",
      "rawMarkdown": "Hi\nI tried to use YOLO on custom data-set but it took very long time on few images. Is it take that much time or there may be some problem from myside?",
      "votes": null
    },
    {
      "id": "687314",
      "postDate": "12/04/2019 08:24:27",
      "content": "<p>which edition do you use? maybe you can try YOLO_v3 rather than YOLO_v1.</p>",
      "rawMarkdown": "which edition do you use? maybe you can try YOLO_v3 rather than YOLO_v1.",
      "votes": null
    },
    {
      "id": "687628",
      "postDate": "12/04/2019 15:27:18",
      "content": "<p>Training the models for images always consume so much time. And,  \"long time\" is not discrete and is relative. So, Its hard to figure out whether problem is on your side or its normal. Like I said training image models usually takes longer. Its better you also try other models also. Like - YOLOv1/v2/v3, RCNN, faster RCNN, etc. and compare the difference with other user's ( browse for articles here, or kdnuggets). This will help you get a clear picture.</p>",
      "rawMarkdown": "Training the models for images always consume so much time. And,  \"long time\" is not discrete and is relative. So, Its hard to figure out whether problem is on your side or its normal. Like I said training image models usually takes longer. Its better you also try other models also. Like - YOLOv1/v2/v3, RCNN, faster RCNN, etc. and compare the difference with other user's ( browse for articles here, or kdnuggets). This will help you get a clear picture.",
      "votes": null
    },
    {
      "id": "687634",
      "postDate": "12/04/2019 15:32:10",
      "content": "<p>👍 </p>",
      "rawMarkdown": "👍",
      "votes": null
    },
    {
      "id": "687661",
      "postDate": "12/04/2019 16:15:41",
      "content": "<p>Try training on one sample, you should be able to overfit one sample almost perfectly and have very low loss. Visualize the predictions from one sample as your train the model and compare to the ground truth. If you your model can't learn one sample then something is wrong. </p>",
      "rawMarkdown": "Try training on one sample, you should be able to overfit one sample almost perfectly and have very low loss. Visualize the predictions from one sample as your train the model and compare to the ground truth. If you your model can't learn one sample then something is wrong.",
      "votes": null
    },
    {
      "id": "688323",
      "postDate": "12/05/2019 13:18:21",
      "content": "<p>Thanks for sharing.</p>",
      "rawMarkdown": "Thanks for sharing.",
      "votes": null
    },
    {
      "id": "688334",
      "postDate": "12/05/2019 13:38:00",
      "content": "<p>my pleasure.</p>",
      "rawMarkdown": "my pleasure.",
      "votes": null
    },
    {
      "id": "688450",
      "postDate": "12/05/2019 15:50:05",
      "content": "<p>CenterNet👍 </p>",
      "rawMarkdown": "CenterNet👍",
      "votes": null
    },
    {
      "id": "688493",
      "postDate": "12/05/2019 16:30:10",
      "content": "<p>Good choice, I also plan to use it.😂 </p>",
      "rawMarkdown": "Good choice, I also plan to use it.😂",
      "votes": null
    },
    {
      "id": "688555",
      "postDate": "12/05/2019 18:25:09",
      "content": "<p>Which implementation of YOLO is better to use? Darkflow or Darknet?</p>",
      "rawMarkdown": "Which implementation of YOLO is better to use? Darkflow or Darknet?",
      "votes": null
    },
    {
      "id": "688619",
      "postDate": "12/05/2019 20:22:52",
      "content": "<p>Very Useful Thx</p>",
      "rawMarkdown": "Very Useful Thx",
      "votes": null
    },
    {
      "id": "688738",
      "postDate": "12/06/2019 02:11:27",
      "content": "<p>Sorry, I'm not sure, maybe you can try it.</p>",
      "rawMarkdown": "Sorry, I'm not sure, maybe you can try it.",
      "votes": null
    },
    {
      "id": "688739",
      "postDate": "12/06/2019 02:11:56",
      "content": "<p>my pleasure.</p>",
      "rawMarkdown": "my pleasure.",
      "votes": null
    },
    {
      "id": "689117",
      "postDate": "12/06/2019 13:49:46",
      "content": "<p>useful</p>",
      "rawMarkdown": "useful",
      "votes": null
    },
    {
      "id": "689118",
      "postDate": "12/06/2019 13:52:21",
      "content": "<p>thanks.\nwow, you have been 46th, good work👍 </p>",
      "rawMarkdown": "thanks.\nwow, you have been 46th, good work👍",
      "votes": null
    },
    {
      "id": "689319",
      "postDate": "12/06/2019 19:21:55",
      "content": "<p>It's fine. Managing dependencies is such a painful process. Google colab seems to be a far better option. No need to worry about dependencies, just bit more data management and you are on your way.</p>",
      "rawMarkdown": "It's fine. Managing dependencies is such a painful process. Google colab seems to be a far better option. No need to worry about dependencies, just bit more data management and you are on your way.",
      "votes": null
    },
    {
      "id": "689889",
      "postDate": "12/07/2019 16:17:26",
      "content": "<p>Hi, I have some confusion about CenterNet, could you please give me some suggestions? thanks very much.😃  <a href=\"/mashlyn\">@mashlyn</a> \n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120458#latest-688895\">https://www.kaggle.com/c/pku-autonomous-driving/discussion/120458#latest-688895</a></p>",
      "rawMarkdown": "Hi, I have some confusion about CenterNet, could you please give me some suggestions? thanks very much.😃  @mashlyn \nhttps://www.kaggle.com/c/pku-autonomous-driving/discussion/120458#latest-688895",
      "votes": null
    },
    {
      "id": "690325",
      "postDate": "12/08/2019 11:45:38",
      "content": "<p>Thanks for those links</p>",
      "rawMarkdown": "Thanks for those links",
      "votes": null
    },
    {
      "id": "690607",
      "postDate": "12/08/2019 21:12:27",
      "content": "<p>Thanks for sharing, it is very useful.</p>",
      "rawMarkdown": "Thanks for sharing, it is very useful.",
      "votes": null
    },
    {
      "id": "691128",
      "postDate": "12/09/2019 16:47:01",
      "content": "<p>Yeah! It's good.\nIt will helpful to Selection algorithm of ML..</p>",
      "rawMarkdown": "Yeah! It's good.\nIt will helpful to Selection algorithm of ML..",
      "votes": null
    },
    {
      "id": "691539",
      "postDate": "12/10/2019 08:45:57",
      "content": "<p>Thanks for sharing.  It's a good collection</p>",
      "rawMarkdown": "Thanks for sharing.  It's a good collection",
      "votes": null
    },
    {
      "id": "691737",
      "postDate": "12/10/2019 12:54:46",
      "content": "<p>what are the requirements before using YOLO.?</p>",
      "rawMarkdown": "what are the requirements before using YOLO.?",
      "votes": null
    },
    {
      "id": "691743",
      "postDate": "12/10/2019 13:06:12",
      "content": "<p>YOLO is very basic, maybe a little keras or pytorch. <a href=\"/santoshd3\">@santoshd3</a> </p>",
      "rawMarkdown": "YOLO is very basic, maybe a little keras or pytorch. @santoshd3",
      "votes": null
    },
    {
      "id": "692505",
      "postDate": "12/11/2019 10:40:54",
      "content": "<p>Thank you for your advice!</p>",
      "rawMarkdown": "Thank you for your advice!",
      "votes": null
    },
    {
      "id": "692980",
      "postDate": "12/11/2019 23:25:49",
      "content": "<p>Thanks for sharing so many great links and resources. Super helpful for first-timers diving into object-detection!</p>",
      "rawMarkdown": "Thanks for sharing so many great links and resources. Super helpful for first-timers diving into object-detection!",
      "votes": null
    },
    {
      "id": "693228",
      "postDate": "12/12/2019 07:02:33",
      "content": "<p>mmdetection is a very good repo to fork when trying out these models.\nsetting these models up and experimenting is a lot of work but this repo modulates all the object detection codes and simplifies the experiments.</p>\n\n<p>it supports all </p>\n\n<p>YOLO_v3\nSSD-github\nCornerNet\nCenterNet-github\nFCOS\nRPN (Region Proposal Network) + anchor\nFaster R-CNN-github\nMask R-CNN-github</p>\n\n<p><a href=\"https://github.com/open-mmlab/mmdetection\">https://github.com/open-mmlab/mmdetection</a></p>",
      "rawMarkdown": "mmdetection is a very good repo to fork when trying out these models.\nsetting these models up and experimenting is a lot of work but this repo modulates all the object detection codes and simplifies the experiments.\n\nit supports all \n\nYOLO_v3\nSSD-github\nCornerNet\nCenterNet-github\nFCOS\nRPN (Region Proposal Network) + anchor\nFaster R-CNN-github\nMask R-CNN-github\n\nhttps://github.com/open-mmlab/mmdetection",
      "votes": null
    },
    {
      "id": "693231",
      "postDate": "12/12/2019 07:08:49",
      "content": "<p>good job <a href=\"/kyoshioka47\">@kyoshioka47</a> </p>",
      "rawMarkdown": "good job @kyoshioka47",
      "votes": null
    },
    {
      "id": "693576",
      "postDate": "12/12/2019 14:24:28",
      "content": "<p>Thanks for sharing!</p>",
      "rawMarkdown": "Thanks for sharing!",
      "votes": null
    },
    {
      "id": "693582",
      "postDate": "12/12/2019 14:33:11",
      "content": "<p>My pleasure, maybe the algorithm is too much to choose, you can just try CenterNet, It's easy to use, and there are so much notebooks for reference.</p>\n\n<p>Using a more powerful backbone and data augmentation, you will get a not bad score <a href=\"/way2edata\">@way2edata</a> @ all guys</p>",
      "rawMarkdown": "My pleasure, maybe the algorithm is too much to choose, you can just try CenterNet, It's easy to use, and there are so much notebooks for reference.\n\nUsing a more powerful backbone and data augmentation, you will get a not bad score @way2edata @ all guys",
      "votes": null
    },
    {
      "id": "693720",
      "postDate": "12/12/2019 18:06:34",
      "content": "<p>Thanks!  </p>",
      "rawMarkdown": "Thanks!",
      "votes": null
    },
    {
      "id": "699588",
      "postDate": "12/20/2019 16:05:14",
      "content": "<p>Thank you for sharing! So, do we need to first implement stage-1 and then move on to stage-2? Like, one after the other?</p>",
      "rawMarkdown": "Thank you for sharing! So, do we need to first implement stage-1 and then move on to stage-2? Like, one after the other?",
      "votes": null
    },
    {
      "id": "699838",
      "postDate": "12/21/2019 02:28:22",
      "content": "<p>You have no time to learn all.</p>",
      "rawMarkdown": "You have no time to learn all.",
      "votes": null
    },
    {
      "id": "701406",
      "postDate": "12/23/2019 13:11:24",
      "content": "<p>How do I collect this? It's great! 👍 </p>",
      "rawMarkdown": "How do I collect this? It's great! 👍",
      "votes": null
    },
    {
      "id": "703772",
      "postDate": "12/26/2019 15:55:50",
      "content": "<p>Thanks. I really learn a lot from your answers</p>",
      "rawMarkdown": "Thanks. I really learn a lot from your answers",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 686399,
      "author_name": "diegojohnson",
      "author_url": "",
      "post_date": "12/03/2019 05:28:44",
      "content": "<p>If you want more references, see this post. <a href=\"/bibek777\">@bibek777</a> \n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/113895#latest-685345\">https://www.kaggle.com/c/pku-autonomous-driving/discussion/113895#latest-685345</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 686482,
      "author_name": "mobassir",
      "author_url": "",
      "post_date": "12/03/2019 07:53:08",
      "content": "<p>following</p>",
      "votes": null,
      "replies": [
        {
          "id": 686515,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/03/2019 08:39:30",
          "content": "<p>Thanks</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 687170,
      "author_name": "mashlyn",
      "author_url": "",
      "post_date": "12/04/2019 02:17:47",
      "content": "<p>Thanks for sharing. Very useful to     me.❗️</p>",
      "votes": null,
      "replies": [
        {
          "id": 687175,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/04/2019 02:28:29",
          "content": "<p>Wow, you have been 43rd, which algorithm do you choose?🤔 </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 688450,
          "author_name": "mashlyn",
          "author_url": "",
          "post_date": "12/05/2019 15:50:05",
          "content": "<p>CenterNet👍 </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 688493,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/05/2019 16:30:10",
          "content": "<p>Good choice, I also plan to use it.😂 </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 689889,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/07/2019 16:17:26",
          "content": "<p>Hi, I have some confusion about CenterNet, could you please give me some suggestions? thanks very much.😃  <a href=\"/mashlyn\">@mashlyn</a> \n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120458#latest-688895\">https://www.kaggle.com/c/pku-autonomous-driving/discussion/120458#latest-688895</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 687306,
      "author_name": "abhishakvarshney",
      "author_url": "",
      "post_date": "12/04/2019 07:57:36",
      "content": "<p>Hi\nI tried to use YOLO on custom data-set but it took very long time on few images. Is it take that much time or there may be some problem from myside?</p>",
      "votes": null,
      "replies": [
        {
          "id": 687314,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/04/2019 08:24:27",
          "content": "<p>which edition do you use? maybe you can try YOLO_v3 rather than YOLO_v1.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 687628,
          "author_name": "atulanandjha",
          "author_url": "",
          "post_date": "12/04/2019 15:27:18",
          "content": "<p>Training the models for images always consume so much time. And,  \"long time\" is not discrete and is relative. So, Its hard to figure out whether problem is on your side or its normal. Like I said training image models usually takes longer. Its better you also try other models also. Like - YOLOv1/v2/v3, RCNN, faster RCNN, etc. and compare the difference with other user's ( browse for articles here, or kdnuggets). This will help you get a clear picture.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 687634,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/04/2019 15:32:10",
          "content": "<p>👍 </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 687661,
          "author_name": "jackvial",
          "author_url": "",
          "post_date": "12/04/2019 16:15:41",
          "content": "<p>Try training on one sample, you should be able to overfit one sample almost perfectly and have very low loss. Visualize the predictions from one sample as your train the model and compare to the ground truth. If you your model can't learn one sample then something is wrong. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 688323,
      "author_name": "",
      "author_url": "",
      "post_date": "12/05/2019 13:18:21",
      "content": "<p>Thanks for sharing.</p>",
      "votes": null,
      "replies": [
        {
          "id": 688334,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/05/2019 13:38:00",
          "content": "<p>my pleasure.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 688555,
      "author_name": "sorcerersupreme",
      "author_url": "",
      "post_date": "12/05/2019 18:25:09",
      "content": "<p>Which implementation of YOLO is better to use? Darkflow or Darknet?</p>",
      "votes": null,
      "replies": [
        {
          "id": 688738,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/06/2019 02:11:27",
          "content": "<p>Sorry, I'm not sure, maybe you can try it.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 689319,
          "author_name": "sorcerersupreme",
          "author_url": "",
          "post_date": "12/06/2019 19:21:55",
          "content": "<p>It's fine. Managing dependencies is such a painful process. Google colab seems to be a far better option. No need to worry about dependencies, just bit more data management and you are on your way.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 688619,
      "author_name": "akirag",
      "author_url": "",
      "post_date": "12/05/2019 20:22:52",
      "content": "<p>Very Useful Thx</p>",
      "votes": null,
      "replies": [
        {
          "id": 688739,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/06/2019 02:11:56",
          "content": "<p>my pleasure.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 689117,
      "author_name": "xuzhou1993",
      "author_url": "",
      "post_date": "12/06/2019 13:49:46",
      "content": "<p>useful</p>",
      "votes": null,
      "replies": [
        {
          "id": 689118,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/06/2019 13:52:21",
          "content": "<p>thanks.\nwow, you have been 46th, good work👍 </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 690325,
      "author_name": "solomon4",
      "author_url": "",
      "post_date": "12/08/2019 11:45:38",
      "content": "<p>Thanks for those links</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 690607,
      "author_name": "muhammedfatihyigit",
      "author_url": "",
      "post_date": "12/08/2019 21:12:27",
      "content": "<p>Thanks for sharing, it is very useful.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 691128,
      "author_name": "kalyanchak",
      "author_url": "",
      "post_date": "12/09/2019 16:47:01",
      "content": "<p>Yeah! It's good.\nIt will helpful to Selection algorithm of ML..</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 691539,
      "author_name": "swathym",
      "author_url": "",
      "post_date": "12/10/2019 08:45:57",
      "content": "<p>Thanks for sharing.  It's a good collection</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 691737,
      "author_name": "santoshd3",
      "author_url": "",
      "post_date": "12/10/2019 12:54:46",
      "content": "<p>what are the requirements before using YOLO.?</p>",
      "votes": null,
      "replies": [
        {
          "id": 691743,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/10/2019 13:06:12",
          "content": "<p>YOLO is very basic, maybe a little keras or pytorch. <a href=\"/santoshd3\">@santoshd3</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 692505,
      "author_name": "namyongjung",
      "author_url": "",
      "post_date": "12/11/2019 10:40:54",
      "content": "<p>Thank you for your advice!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 692980,
      "author_name": "teejmahal20",
      "author_url": "",
      "post_date": "12/11/2019 23:25:49",
      "content": "<p>Thanks for sharing so many great links and resources. Super helpful for first-timers diving into object-detection!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 693228,
      "author_name": "kyoshioka47",
      "author_url": "",
      "post_date": "12/12/2019 07:02:33",
      "content": "<p>mmdetection is a very good repo to fork when trying out these models.\nsetting these models up and experimenting is a lot of work but this repo modulates all the object detection codes and simplifies the experiments.</p>\n\n<p>it supports all </p>\n\n<p>YOLO_v3\nSSD-github\nCornerNet\nCenterNet-github\nFCOS\nRPN (Region Proposal Network) + anchor\nFaster R-CNN-github\nMask R-CNN-github</p>\n\n<p><a href=\"https://github.com/open-mmlab/mmdetection\">https://github.com/open-mmlab/mmdetection</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 693231,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/12/2019 07:08:49",
          "content": "<p>good job <a href=\"/kyoshioka47\">@kyoshioka47</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 693576,
      "author_name": "way2edata",
      "author_url": "",
      "post_date": "12/12/2019 14:24:28",
      "content": "<p>Thanks for sharing!</p>",
      "votes": null,
      "replies": [
        {
          "id": 693582,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/12/2019 14:33:11",
          "content": "<p>My pleasure, maybe the algorithm is too much to choose, you can just try CenterNet, It's easy to use, and there are so much notebooks for reference.</p>\n\n<p>Using a more powerful backbone and data augmentation, you will get a not bad score <a href=\"/way2edata\">@way2edata</a> @ all guys</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 693720,
      "author_name": "takabon",
      "author_url": "",
      "post_date": "12/12/2019 18:06:34",
      "content": "<p>Thanks!  </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 699588,
      "author_name": "panivishal",
      "author_url": "",
      "post_date": "12/20/2019 16:05:14",
      "content": "<p>Thank you for sharing! So, do we need to first implement stage-1 and then move on to stage-2? Like, one after the other?</p>",
      "votes": null,
      "replies": [
        {
          "id": 699838,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/21/2019 02:28:22",
          "content": "<p>You have no time to learn all.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 701406,
      "author_name": "p90rushx",
      "author_url": "",
      "post_date": "12/23/2019 13:11:24",
      "content": "<p>How do I collect this? It's great! 👍 </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 703772,
      "author_name": "",
      "author_url": "",
      "post_date": "12/26/2019 15:55:50",
      "content": "<p>Thanks. I really learn a lot from your answers</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "686395": "### Upvote if you like, thanks. \n\nThis is a big picture of **object detection**.\n\nThe algorithm of object detection is devided into two types - **1-stage** and **2-stage**.\n\n## 1-stage\n- anchor_based\n    * [YOLO_v3](https://pjreddie.com/media/files/papers/YOLOv3.pdf)\n    * [SSD](http://arxiv.org/abs/1512.02325)-[github](https://github.com/amdegroot/ssd.pytorch)\n- anchor-free\n    * [CornerNet](https://arxiv.org/abs/1808.01244)\n    * [CenterNet](https://arxiv.org/pdf/1904.07850.pdf)-[github](https://github.com/xingyizhou/CenterNet)\n    * [FCOS](https://arxiv.org/pdf/1904.01355.pdf)\n\n## 2-stage\n-  RPN (Region Proposal Network) + anchor\n    * [Faster R-CNN](https://arxiv.org/abs/1506.01497)-[github](https://github.com/endernewton/tf-faster-rcnn)\n    * [Mask R-CNN](https://arxiv.org/abs/1703.06870)-[github](https://github.com/facebookresearch/Detectron)\n\nI just list some basic algorithm, you can google it to learn more.\n\n### Finally, I'll give you some references which are basically same task with this competition.\n\n* [SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again](https://arxiv.org/pdf/1711.10006.pdf)\n[Github link] (https://github.com/wadimkehl/ssd-6d)\n* [Bag of Freebies for Training Object Detection Neural Networks](https://arxiv.org/abs/1902.04103)\n* [DeepIM: Deep Iterative Matching for 6D Pose Estimation](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120510)\n\nThis post will be updated as I find new useful references.\n\n### This is my other topics, hope you like.\n[The coordinate system of this competition](https://www.kaggle.com/c/pku-autonomous-driving/discussion/123385)\n[The algorithm that baidu apollo chooses!](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120076)\n[Understant Camera Intrinsic Parameters and X, Y, Z](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120083#latest-686785)\n[Algorithm Selection in 6D Pose Estimation](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120381#latest-688477)\n[How the Competition Organizer get train.csv !](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120443)\n[A way to improve the accuracy of model](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120653)\n[Object Detection in 20 Years](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120710)\n\n\nMy notebook:\n[Centernet - Obejects as Points](https://www.kaggle.com/diegojohnson/centernet-objects-as-points)\n[A Way to Regress Translation and Rotation](https://www.kaggle.com/diegojohnson/a-way-to-regress-translation-and-rotation)\n[EfficientDet-D1](https://www.kaggle.com/diegojohnson/best-algorithm-so-far-google-s-new-paper-in-nov)\n[Regress angle in 2 bins - centernet way](https://www.kaggle.com/diegojohnson/a-better-way-to-regress-yaw)\n[A clear view of car pose](https://www.kaggle.com/diegojohnson/a-clear-view-of-car-pose)\n[Dataset with preprocess](https://www.kaggle.com/c/pku-autonomous-driving/discussion/124480)",
    "686399": "If you want more references, see this post. @bibek777 \nhttps://www.kaggle.com/c/pku-autonomous-driving/discussion/113895#latest-685345",
    "686482": "following",
    "686515": "Thanks",
    "687170": "Thanks for sharing. Very useful to     me.❗️",
    "687175": "Wow, you have been 43rd, which algorithm do you choose?🤔",
    "687306": "Hi\nI tried to use YOLO on custom data-set but it took very long time on few images. Is it take that much time or there may be some problem from myside?",
    "687314": "which edition do you use? maybe you can try YOLO_v3 rather than YOLO_v1.",
    "687628": "Training the models for images always consume so much time. And,  \"long time\" is not discrete and is relative. So, Its hard to figure out whether problem is on your side or its normal. Like I said training image models usually takes longer. Its better you also try other models also. Like - YOLOv1/v2/v3, RCNN, faster RCNN, etc. and compare the difference with other user's ( browse for articles here, or kdnuggets). This will help you get a clear picture.",
    "687634": "👍",
    "687661": "Try training on one sample, you should be able to overfit one sample almost perfectly and have very low loss. Visualize the predictions from one sample as your train the model and compare to the ground truth. If you your model can't learn one sample then something is wrong.",
    "688323": "Thanks for sharing.",
    "688334": "my pleasure.",
    "688450": "CenterNet👍",
    "688493": "Good choice, I also plan to use it.😂",
    "688555": "Which implementation of YOLO is better to use? Darkflow or Darknet?",
    "688619": "Very Useful Thx",
    "688738": "Sorry, I'm not sure, maybe you can try it.",
    "688739": "my pleasure.",
    "689117": "useful",
    "689118": "thanks.\nwow, you have been 46th, good work👍",
    "689319": "It's fine. Managing dependencies is such a painful process. Google colab seems to be a far better option. No need to worry about dependencies, just bit more data management and you are on your way.",
    "689889": "Hi, I have some confusion about CenterNet, could you please give me some suggestions? thanks very much.😃  @mashlyn \nhttps://www.kaggle.com/c/pku-autonomous-driving/discussion/120458#latest-688895",
    "690325": "Thanks for those links",
    "690607": "Thanks for sharing, it is very useful.",
    "691128": "Yeah! It's good.\nIt will helpful to Selection algorithm of ML..",
    "691539": "Thanks for sharing.  It's a good collection",
    "691737": "what are the requirements before using YOLO.?",
    "691743": "YOLO is very basic, maybe a little keras or pytorch. @santoshd3",
    "692505": "Thank you for your advice!",
    "692980": "Thanks for sharing so many great links and resources. Super helpful for first-timers diving into object-detection!",
    "693228": "mmdetection is a very good repo to fork when trying out these models.\nsetting these models up and experimenting is a lot of work but this repo modulates all the object detection codes and simplifies the experiments.\n\nit supports all \n\nYOLO_v3\nSSD-github\nCornerNet\nCenterNet-github\nFCOS\nRPN (Region Proposal Network) + anchor\nFaster R-CNN-github\nMask R-CNN-github\n\nhttps://github.com/open-mmlab/mmdetection",
    "693231": "good job @kyoshioka47",
    "693576": "Thanks for sharing!",
    "693582": "My pleasure, maybe the algorithm is too much to choose, you can just try CenterNet, It's easy to use, and there are so much notebooks for reference.\n\nUsing a more powerful backbone and data augmentation, you will get a not bad score @way2edata @ all guys",
    "693720": "Thanks!",
    "699588": "Thank you for sharing! So, do we need to first implement stage-1 and then move on to stage-2? Like, one after the other?",
    "699838": "You have no time to learn all.",
    "701406": "How do I collect this? It's great! 👍",
    "703772": "Thanks. I really learn a lot from your answers"
  },
  "source": "meta"
}