{ "id": 123683, "title": "Ground Plane Polling", "url": "/competitions/pku-autonomous-driving/discussion/123683", "author_name": "GreatGameDota", "post_date": "2019-12-29T15:47:46.688000", "votes": 7, "comment_count": 0, "views": 0, "content": "

For awhile I was looking at every paper in the 3D Bounding box portion of KITTI benchmark. I would use the pretrained model and predict on this dataset and see if they got good results. For the most part the results were terrible but I found a pretty unknown paper and model that actually worked really well on this dataset.\n\"\"\nThe model returns angles (in euler but different axis/offsets), 3d locations, residuals, scores, classes, 2d boxes, 3d box dimensions, keypoints, and orientation class.
\n\"\"\nI have experimented with this model for a bit and haven't been able to get accurate predictions so I wanted to share. If you are interested in testing this I can give additional info on how I predicted the above image (backbone, calibration).\nCode: https://github.com/arangesh/Ground-Plane-Polling
\nPaper: https://arxiv.org/pdf/1811.06666.pdf

", "messages": [ { "id": 705870, "postDate": "2019-12-29T15:47:46.687Z", "content": "

For awhile I was looking at every paper in the 3D Bounding box portion of KITTI benchmark. I would use the pretrained model and predict on this dataset and see if they got good results. For the most part the results were terrible but I found a pretty unknown paper and model that actually worked really well on this dataset.\n\"\"\nThe model returns angles (in euler but different axis/offsets), 3d locations, residuals, scores, classes, 2d boxes, 3d box dimensions, keypoints, and orientation class.
\n\"\"\nI have experimented with this model for a bit and haven't been able to get accurate predictions so I wanted to share. If you are interested in testing this I can give additional info on how I predicted the above image (backbone, calibration).\nCode: https://github.com/arangesh/Ground-Plane-Polling
\nPaper: https://arxiv.org/pdf/1811.06666.pdf

", "rawMarkdown": "For awhile I was looking at every paper in the [3D Bounding box portion of KITTI benchmark](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d). I would use the pretrained model and predict on this dataset and see if they got good results. For the most part the results were terrible but I found a pretty unknown paper and model that actually worked really well on this dataset.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3543139%2Fc6039028bd7e1a7ab6ab49317d72124c%2FID_06275a4b2.jpg?generation=1577633428865771&alt=media)\nThe model returns angles (in euler but different axis/offsets), 3d locations, residuals, scores, classes, 2d boxes, 3d box dimensions, keypoints, and orientation class. \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3543139%2Ff33357e86b82be4270be79edb481deeb%2Fede4f4c50beaaa1ed4586a9a906c664c_waifu2x_art_noise1_scale_tta_1.png?generation=1577634440816287&alt=media)\nI have experimented with this model for a bit and haven't been able to get accurate predictions so I wanted to share. If you are interested in testing this I can give additional info on how I predicted the above image (backbone, calibration).\nCode: https://github.com/arangesh/Ground-Plane-Polling \nPaper: https://arxiv.org/pdf/1811.06666.pdf", "votes": 7 } ], "comments": [], "raw_markdown_by_id": { "705870": "For awhile I was looking at every paper in the [3D Bounding box portion of KITTI benchmark](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d). I would use the pretrained model and predict on this dataset and see if they got good results. For the most part the results were terrible but I found a pretty unknown paper and model that actually worked really well on this dataset.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3543139%2Fc6039028bd7e1a7ab6ab49317d72124c%2FID_06275a4b2.jpg?generation=1577633428865771&alt=media)\nThe model returns angles (in euler but different axis/offsets), 3d locations, residuals, scores, classes, 2d boxes, 3d box dimensions, keypoints, and orientation class. \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3543139%2Ff33357e86b82be4270be79edb481deeb%2Fede4f4c50beaaa1ed4586a9a906c664c_waifu2x_art_noise1_scale_tta_1.png?generation=1577634440816287&alt=media)\nI have experimented with this model for a bit and haven't been able to get accurate predictions so I wanted to share. If you are interested in testing this I can give additional info on how I predicted the above image (backbone, calibration).\nCode: https://github.com/arangesh/Ground-Plane-Polling \nPaper: https://arxiv.org/pdf/1811.06666.pdf" } }