![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F761152%2Fadcfa4df972dfe71bde7404745460bb1%2Fpic-selected-191130-2044-46.png?generation=1575126996055584&alt=media\")
Hey guys,
\n\nFirst of all, I'd like to thank the organizers for this awesome challenge. This was not a typical computer vision competition that's for sure and I spent a good amount of time just to understand the problem and the data. I went through multiple papers related to 3d object detection. I chose to go with SECOND:
\n\n
I tweaked the SECOND's original source code to get it working for lyft dataset. Added the evaluation metric of the competition for val set evaluation. Modified the default config files according to lyft object dimensions (anchor_ranges, sizes etc). Splitted train/val based on scenes (80:20), configs/nuscenes/all.pp.lowa.config with VoxelNet and PillarFeatureNet worked best for me.
check out my source code: https://github.com/pyaf/second.pytorch
", "messages": [ { "id": 674330, "postDate": "2019-11-16T09:31:34.880Z", "content": "Hey guys,
\n\nFirst of all, I'd like to thank the organizers for this awesome challenge. This was not a typical computer vision competition that's for sure and I spent a good amount of time just to understand the problem and the data. I went through multiple papers related to 3d object detection. I chose to go with SECOND:
\n\n
I tweaked the SECOND's original source code to get it working for lyft dataset. Added the evaluation metric of the competition for val set evaluation. Modified the default config files according to lyft object dimensions (anchor_ranges, sizes etc). Splitted train/val based on scenes (80:20), configs/nuscenes/all.pp.lowa.config with VoxelNet and PillarFeatureNet worked best for me.
check out my source code: https://github.com/pyaf/second.pytorch
", "rawMarkdown": "Hey guys,\n\nFirst of all, I'd like to thank the organizers for this awesome challenge. This was not a typical computer vision competition that's for sure and I spent a good amount of time just to understand the problem and the data. I went through multiple papers related to 3d object detection. I chose to go with [SECOND](https://github.com/traveller59/second.pytorch):\n\n\n\nI tweaked the SECOND's original source code to get it working for lyft dataset. Added the evaluation metric of the competition for val set evaluation. Modified the default config files according to lyft object dimensions (`anchor_ranges`, `sizes` etc). Splitted train/val based on scenes (80:20), [configs/nuscenes/all.pp.lowa.config](https://github.com/pyaf/second.pytorch/blob/master/second/configs/nuscenes/all.pp.lowa.config) with VoxelNet and PillarFeatureNet worked best for me. \n\n\ncheck out my source code: https://github.com/pyaf/second.pytorch ", "votes": 24 }, { "id": 684191, "postDate": "2019-11-29T11:32:20.337Z", "content": "Congrats. Nice work!
", "rawMarkdown": "Congrats. Nice work!", "votes": 1, "replies": [ { "id": 684221, "postDate": "2019-11-29T12:34:47.130Z", "content": "Thanks @auxlchen
", "rawMarkdown": "Thanks @auxlchen " } ] }, { "id": 674536, "postDate": "2019-11-16T17:07:46.047Z", "content": "Great work @rishabhiitbhu! I also used voxel_size : [0.25, 0.25, 20], it worked surprisingly well considering it is much taller than the the voxel size of [0.2, 0.2, 0.4] from the original VoxelNet paper. Still it is something I would have liked to try tuning but I ran out of time.
Sorry to trouble you. It seems that the Lyft dataset link in your github is lost. Could you please update it in your github? Thanks a lot.
", "rawMarkdown": "Sorry to trouble you. It seems that the Lyft dataset link in your github is lost. Could you please update it in your github? Thanks a lot." }, { "id": 678135, "postDate": "2019-11-21T02:28:28.723Z", "content": "Thanks for sharing @rishabhiitbhu! I was trying to follow your codes and to train a model, but while evaluation I can't find file \"gt_data_val.json\", may I ask which file that you used to generate this file? Thanks!!
", "rawMarkdown": "Thanks for sharing @rishabhiitbhu! I was trying to follow your codes and to train a model, but while evaluation I can't find file \"gt_data_val.json\", may I ask which file that you used to generate this file? Thanks!!", "replies": [ { "id": 678262, "postDate": "2019-11-21T07:01:24.117Z", "content": "Checkout prepare.ipynb in the notebooks folder
", "rawMarkdown": "Checkout prepare.ipynb in the notebooks folder", "votes": 2 }, { "id": 678548, "postDate": "2019-11-21T14:20:20.140Z", "content": "Gotcha! Thanks!
", "rawMarkdown": "Gotcha! Thanks!" } ] }, { "id": 678126, "postDate": "2019-11-21T02:18:06.017Z", "content": "Congrats.\nThanks for sharing your code and all your sharing throughout the competition.
", "rawMarkdown": "Congrats.\nThanks for sharing your code and all your sharing throughout the competition. " }, { "id": 674514, "postDate": "2019-11-16T16:03:56.123Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 684191, "author_name": "Xuliang Chen", "author_url": "", "post_date": "2019-11-29T11:32:20.337000", "content": "Congrats. Nice work!
", "votes": 1, "replies": [ { "id": 684221, "author_name": "Rishabh Agrahari", "author_url": "", "post_date": "2019-11-29T12:34:47.130000", "content": "Thanks @auxlchen
", "votes": 0, "replies": [] } ] }, { "id": 674536, "author_name": "Jack Vial", "author_url": "", "post_date": "2019-11-16T17:07:46.047000", "content": "Great work @rishabhiitbhu! I also used voxel_size : [0.25, 0.25, 20], it worked surprisingly well considering it is much taller than the the voxel size of [0.2, 0.2, 0.4] from the original VoxelNet paper. Still it is something I would have liked to try tuning but I ran out of time.
Sorry to trouble you. It seems that the Lyft dataset link in your github is lost. Could you please update it in your github? Thanks a lot.
", "votes": 0, "replies": [] }, { "id": 678135, "author_name": "jionie", "author_url": "", "post_date": "2019-11-21T02:28:28.723000", "content": "Thanks for sharing @rishabhiitbhu! I was trying to follow your codes and to train a model, but while evaluation I can't find file \"gt_data_val.json\", may I ask which file that you used to generate this file? Thanks!!
", "votes": 0, "replies": [ { "id": 678262, "author_name": "Rishabh Agrahari", "author_url": "", "post_date": "2019-11-21T07:01:24.117000", "content": "Checkout prepare.ipynb in the notebooks folder
", "votes": 2, "replies": [] }, { "id": 678548, "author_name": "jionie", "author_url": "", "post_date": "2019-11-21T14:20:20.140000", "content": "Gotcha! Thanks!
", "votes": 0, "replies": [] } ] }, { "id": 678126, "author_name": "Bob Wang", "author_url": "", "post_date": "2019-11-21T02:18:06.017000", "content": "Congrats.\nThanks for sharing your code and all your sharing throughout the competition.
", "votes": 0, "replies": [] }, { "id": 674514, "author_name": "", "author_url": "", "post_date": "2019-11-16T16:03:56.123000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "674330": "Hey guys,\n\nFirst of all, I'd like to thank the organizers for this awesome challenge. This was not a typical computer vision competition that's for sure and I spent a good amount of time just to understand the problem and the data. I went through multiple papers related to 3d object detection. I chose to go with [SECOND](https://github.com/traveller59/second.pytorch):\n\n\n\nI tweaked the SECOND's original source code to get it working for lyft dataset. Added the evaluation metric of the competition for val set evaluation. Modified the default config files according to lyft object dimensions (`anchor_ranges`, `sizes` etc). Splitted train/val based on scenes (80:20), [configs/nuscenes/all.pp.lowa.config](https://github.com/pyaf/second.pytorch/blob/master/second/configs/nuscenes/all.pp.lowa.config) with VoxelNet and PillarFeatureNet worked best for me. \n\n\ncheck out my source code: https://github.com/pyaf/second.pytorch ", "684191": "Congrats. Nice work!", "674536": "Great work @rishabhiitbhu! I also used `voxel_size : [0.25, 0.25, 20]`, it worked surprisingly well considering it is much taller than the the voxel size of `[0.2, 0.2, 0.4]` from the original VoxelNet paper. Still it is something I would have liked to try tuning but I ran out of time. ", "1324447": "Sorry to trouble you. It seems that the Lyft dataset link in your github is lost. Could you please update it in your github? Thanks a lot.", "678135": "Thanks for sharing @rishabhiitbhu! I was trying to follow your codes and to train a model, but while evaluation I can't find file \"gt_data_val.json\", may I ask which file that you used to generate this file? Thanks!!", "678126": "Congrats.\nThanks for sharing your code and all your sharing throughout the competition. ", "674514": "" } }