At first I'd like to express my appreciation for my teammate @phalanx @hesene @lanjunyelan, you guys are awesome! And congrats for those who finished with prize or gold zone:) I also would like to thank to kaggle and host for hosting such a nice competition!
\n\nI wrote a brief summary of my solution and opened up the code in my github.
\n\nOur entire solution is in this thread, please check it out.
\n\nI started with hocop1's really great kernel, so basic part of my code is based on his.\nMy main contributions are below:\n - Change the CenterNet implementation to see---'s code for kuzushiji competition\n - Change backbone to resnet152\n - change the depth representation to 1 / (sigmoid(z) - 1\n - remove x, y, yaw(actually pitch), roll prediction\n - remove optimize xyz, just convert rcz to xyz\n - increase image size to (832, 2080) or (960, 2400) or (1088, 2720)\n - change the distance threshold to 4\n - change the confidence threshold to 0.1\n - ensemble 5 folds\n - remove false positive by using masks in 'test_masks/'
\n\nI tried 3 scale ensemble, but it might be not so different from just 5 fold ensemble of image_size=(1088,2720).
\n My single fold model achieve public0.097/private0.091 and 5 fold ensemble got public0.117/private0.105.
\n With ensmebling Phalanx's model and Jhui's model, we got public0.129/private0.119.
code is available here.
", "messages": [ { "id": "725356", "postDate": "01/22/2020 02:47:02", "content": "At first I'd like to express my appreciation for my teammate @phalanx @hesene @lanjunyelan, you guys are awesome! And congrats for those who finished with prize or gold zone:) I also would like to thank to kaggle and host for hosting such a nice competition!
\n\nI wrote a brief summary of my solution and opened up the code in my github.
\n\nOur entire solution is in this thread, please check it out.
\n\nI started with hocop1's really great kernel, so basic part of my code is based on his.\nMy main contributions are below:\n - Change the CenterNet implementation to see---'s code for kuzushiji competition\n - Change backbone to resnet152\n - change the depth representation to 1 / (sigmoid(z) - 1\n - remove x, y, yaw(actually pitch), roll prediction\n - remove optimize xyz, just convert rcz to xyz\n - increase image size to (832, 2080) or (960, 2400) or (1088, 2720)\n - change the distance threshold to 4\n - change the confidence threshold to 0.1\n - ensemble 5 folds\n - remove false positive by using masks in 'test_masks/'
\n\nI tried 3 scale ensemble, but it might be not so different from just 5 fold ensemble of image_size=(1088,2720).
\n My single fold model achieve public0.097/private0.091 and 5 fold ensemble got public0.117/private0.105.
\n With ensmebling Phalanx's model and Jhui's model, we got public0.129/private0.119.
code is available here.
", "rawMarkdown": "At first I'd like to express my appreciation for my teammate @phalanx @hesene @lanjunyelan, you guys are awesome! And congrats for those who finished with prize or gold zone:) I also would like to thank to kaggle and host for hosting such a nice competition!\n\nI wrote a brief summary of my solution and opened up the code in [my github](https://github.com/bamps53/kaggle-autonomous-driving2019).\n\nOur entire solution is in [this thread](https://www.kaggle.com/c/pku-autonomous-driving/discussion/127034), please check it out. \n\n---\n\nI started with [hocop1's really great kernel](https://www.kaggle.com/hocop1/centernet-baseline), so basic part of my code is based on his.\nMy main contributions are below:\n - Change the CenterNet implementation to [see---'s code for kuzushiji competition](https://github.com/see--/kuzushiji-recognition)\n - Change backbone to resnet152\n - change the depth representation to 1 / (sigmoid(z) - 1\n - remove x, y, yaw(actually pitch), roll prediction\n - remove optimize xyz, just convert rcz to xyz\n - increase image size to (832, 2080) or (960, 2400) or (1088, 2720)\n - change the distance threshold to 4\n - change the confidence threshold to 0.1\n - ensemble 5 folds\n - remove false positive by using masks in 'test_masks/'\n \n\nI tried 3 scale ensemble, but it might be not so different from just 5 fold ensemble of image_size=(1088,2720). \n My single fold model achieve public0.097/private0.091 and 5 fold ensemble got public0.117/private0.105. \n With ensmebling Phalanx's model and Jhui's model, we got public0.129/private0.119. \n\n---\ncode is available [here](https://github.com/bamps53/kaggle-autonomous-driving2019).", "votes": null }, { "id": "725402", "postDate": "01/22/2020 03:52:22", "content": "Awesome solution,I learned many from it.Thanks for sharing codes.I can't understand why you remove optimize xyz, just convert rcz to xyz.
", "rawMarkdown": "Awesome solution,I learned many from it.Thanks for sharing codes.I can't understand why you remove optimize xyz, just convert rcz to xyz.", "votes": null }, { "id": "725616", "postDate": "01/22/2020 09:52:35", "content": "Congrats dude...\n1) Could throw some light on what is it trying to do..
\n\n```\ndef gather_embeddings(self, embeddings, centers):\n gathered_embeddings = []\n for sample_index in range(len(centers)):\n center_mask = centers[sample_index, :, 0] != -1\n if center_mask.sum().item() == 0:\n continue\n per_sample_centers = centers[sample_index][center_mask]\n emb = embeddings[sample_index][:, per_sample_centers[:, 1],\n per_sample_centers[:, 0]].transpose(0, 1)\n gathered_embeddings.append(emb)\n gathered_embeddings = torch.cat(gathered_embeddings, 0)
\n\n return gathered_embeddings\n```
\n\n2)\" depth representation to 1 / (sigmoid(z) - 1 \" what is motivation for this in Loss function where i think u are using this.
", "rawMarkdown": "Congrats dude...\n1) Could throw some light on what is it trying to do..\n\n \n```\ndef gather_embeddings(self, embeddings, centers):\n gathered_embeddings = []\n for sample_index in range(len(centers)):\n center_mask = centers[sample_index, :, 0] != -1\n if center_mask.sum().item() == 0:\n continue\n per_sample_centers = centers[sample_index][center_mask]\n emb = embeddings[sample_index][:, per_sample_centers[:, 1],\n per_sample_centers[:, 0]].transpose(0, 1)\n gathered_embeddings.append(emb)\n gathered_embeddings = torch.cat(gathered_embeddings, 0)\n\n return gathered_embeddings\n\n```\n \n\n2)\" depth representation to 1 / (sigmoid(z) - 1 \" what is motivation for this in Loss function where i think u are using this.", "votes": null }, { "id": "725850", "postDate": "01/22/2020 14:46:02", "content": "Hi @ynhuhu, I tried both and score was almost same or slightly better if you just convert. Somehow x,y predictions were not so accurate and were just noise for me. I also tried to optimize distance in world coordinate directly, but just converting was best. And also optimize xyz was so slow.
", "rawMarkdown": "Hi @ynhuhu, I tried both and score was almost same or slightly better if you just convert. Somehow x,y predictions were not so accurate and were just noise for me. I also tried to optimize distance in world coordinate directly, but just converting was best. And also optimize xyz was so slow.", "votes": null }, { "id": "725859", "postDate": "01/22/2020 14:52:12", "content": "Hi @jaideepvalani,\n1) I just copied code from see---'s implementation and I haven't used this function for my solution. But what this function does is collect prediction for the point at where the ground truth is.\n2) I just followed the way CenterNet original paper or other 3d monocular prediciton paper use. I guess it helps model to represent wide range of numerical value,but some participants report it didn't work for him, so it could be not necessary.
", "rawMarkdown": "Hi @jaideepvalani,\n1) I just copied code from see---'s implementation and I haven't used this function for my solution. But what this function does is collect prediction for the point at where the ground truth is.\n2) I just followed the way CenterNet original paper or other 3d monocular prediciton paper use. I guess it helps model to represent wide range of numerical value,but some participants report it didn't work for him, so it could be not necessary.", "votes": null }, { "id": "727498", "postDate": "01/23/2020 18:47:40", "content": "@bamps53 thanks for the explanations. could you tell how long was the training (1 fold)? Thanks!
", "rawMarkdown": "bamps53 thanks for the explanations. could you tell how long was the training (1 fold)? Thanks!", "votes": null }, { "id": "727689", "postDate": "01/24/2020 01:05:55", "content": "For largest image size=(1088, 2720), it took 1 and half days to train 1 fold with 2080Ti....
", "rawMarkdown": "For largest image size=(1088, 2720), it took 1 and half days to train 1 fold with 2080Ti....", "votes": null }, { "id": "732538", "postDate": "01/29/2020 23:26:55", "content": "Congrats and thanks for sharing!
", "rawMarkdown": "Congrats and thanks for sharing!", "votes": null } ], "comments": [ { "id": 725402, "author_name": "ynhuhu", "author_url": "", "post_date": "01/22/2020 03:52:22", "content": "Awesome solution,I learned many from it.Thanks for sharing codes.I can't understand why you remove optimize xyz, just convert rcz to xyz.
", "votes": null, "replies": [ { "id": 725850, "author_name": "bamps53", "author_url": "", "post_date": "01/22/2020 14:46:02", "content": "Hi @ynhuhu, I tried both and score was almost same or slightly better if you just convert. Somehow x,y predictions were not so accurate and were just noise for me. I also tried to optimize distance in world coordinate directly, but just converting was best. And also optimize xyz was so slow.
", "votes": null, "replies": [] } ] }, { "id": 725616, "author_name": "jaideepvalani", "author_url": "", "post_date": "01/22/2020 09:52:35", "content": "Congrats dude...\n1) Could throw some light on what is it trying to do..
\n\n```\ndef gather_embeddings(self, embeddings, centers):\n gathered_embeddings = []\n for sample_index in range(len(centers)):\n center_mask = centers[sample_index, :, 0] != -1\n if center_mask.sum().item() == 0:\n continue\n per_sample_centers = centers[sample_index][center_mask]\n emb = embeddings[sample_index][:, per_sample_centers[:, 1],\n per_sample_centers[:, 0]].transpose(0, 1)\n gathered_embeddings.append(emb)\n gathered_embeddings = torch.cat(gathered_embeddings, 0)
\n\n return gathered_embeddings\n```
\n\n2)\" depth representation to 1 / (sigmoid(z) - 1 \" what is motivation for this in Loss function where i think u are using this.
", "votes": null, "replies": [ { "id": 725859, "author_name": "bamps53", "author_url": "", "post_date": "01/22/2020 14:52:12", "content": "Hi @jaideepvalani,\n1) I just copied code from see---'s implementation and I haven't used this function for my solution. But what this function does is collect prediction for the point at where the ground truth is.\n2) I just followed the way CenterNet original paper or other 3d monocular prediciton paper use. I guess it helps model to represent wide range of numerical value,but some participants report it didn't work for him, so it could be not necessary.
", "votes": null, "replies": [] } ] }, { "id": 727498, "author_name": "jesucristo", "author_url": "", "post_date": "01/23/2020 18:47:40", "content": "@bamps53 thanks for the explanations. could you tell how long was the training (1 fold)? Thanks!
", "votes": null, "replies": [ { "id": 727689, "author_name": "bamps53", "author_url": "", "post_date": "01/24/2020 01:05:55", "content": "For largest image size=(1088, 2720), it took 1 and half days to train 1 fold with 2080Ti....
", "votes": null, "replies": [] } ] }, { "id": 732538, "author_name": "corochann", "author_url": "", "post_date": "01/29/2020 23:26:55", "content": "Congrats and thanks for sharing!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "725356": "At first I'd like to express my appreciation for my teammate @phalanx @hesene @lanjunyelan, you guys are awesome! And congrats for those who finished with prize or gold zone:) I also would like to thank to kaggle and host for hosting such a nice competition!\n\nI wrote a brief summary of my solution and opened up the code in [my github](https://github.com/bamps53/kaggle-autonomous-driving2019).\n\nOur entire solution is in [this thread](https://www.kaggle.com/c/pku-autonomous-driving/discussion/127034), please check it out. \n\n---\n\nI started with [hocop1's really great kernel](https://www.kaggle.com/hocop1/centernet-baseline), so basic part of my code is based on his.\nMy main contributions are below:\n - Change the CenterNet implementation to [see---'s code for kuzushiji competition](https://github.com/see--/kuzushiji-recognition)\n - Change backbone to resnet152\n - change the depth representation to 1 / (sigmoid(z) - 1\n - remove x, y, yaw(actually pitch), roll prediction\n - remove optimize xyz, just convert rcz to xyz\n - increase image size to (832, 2080) or (960, 2400) or (1088, 2720)\n - change the distance threshold to 4\n - change the confidence threshold to 0.1\n - ensemble 5 folds\n - remove false positive by using masks in 'test_masks/'\n \n\nI tried 3 scale ensemble, but it might be not so different from just 5 fold ensemble of image_size=(1088,2720). \n My single fold model achieve public0.097/private0.091 and 5 fold ensemble got public0.117/private0.105. \n With ensmebling Phalanx's model and Jhui's model, we got public0.129/private0.119. \n\n---\ncode is available [here](https://github.com/bamps53/kaggle-autonomous-driving2019).", "725402": "Awesome solution,I learned many from it.Thanks for sharing codes.I can't understand why you remove optimize xyz, just convert rcz to xyz.", "725616": "Congrats dude...\n1) Could throw some light on what is it trying to do..\n\n \n```\ndef gather_embeddings(self, embeddings, centers):\n gathered_embeddings = []\n for sample_index in range(len(centers)):\n center_mask = centers[sample_index, :, 0] != -1\n if center_mask.sum().item() == 0:\n continue\n per_sample_centers = centers[sample_index][center_mask]\n emb = embeddings[sample_index][:, per_sample_centers[:, 1],\n per_sample_centers[:, 0]].transpose(0, 1)\n gathered_embeddings.append(emb)\n gathered_embeddings = torch.cat(gathered_embeddings, 0)\n\n return gathered_embeddings\n\n```\n \n\n2)\" depth representation to 1 / (sigmoid(z) - 1 \" what is motivation for this in Loss function where i think u are using this.", "725850": "Hi @ynhuhu, I tried both and score was almost same or slightly better if you just convert. Somehow x,y predictions were not so accurate and were just noise for me. I also tried to optimize distance in world coordinate directly, but just converting was best. And also optimize xyz was so slow.", "725859": "Hi @jaideepvalani,\n1) I just copied code from see---'s implementation and I haven't used this function for my solution. But what this function does is collect prediction for the point at where the ground truth is.\n2) I just followed the way CenterNet original paper or other 3d monocular prediciton paper use. I guess it helps model to represent wide range of numerical value,but some participants report it didn't work for him, so it could be not necessary.", "727498": "bamps53 thanks for the explanations. could you tell how long was the training (1 fold)? Thanks!", "727689": "For largest image size=(1088, 2720), it took 1 and half days to train 1 fold with 2080Ti....", "732538": "Congrats and thanks for sharing!" }, "source": "meta" }