search

yifanzhang713 / IA-SSD

Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds (CVPR 2022, Oral)
Apache License 2.0
383 stars 50 forks source link

using the code to train model on KITTI dataset, but get only 78.5% mAP on Car Moderate #54

Open lubinBoooos opened 1 year ago

lubinBoooos commented 1 year ago

I git the code, and train it with 4 GPUs, the final validation result is below, and the result match the author's log: Car AP@0.70, 0.70, 0.70: bbox AP:96.1959, 90.2071, 89.5205 bev AP:90.3205, 88.6980, 87.1650 3d AP:89.3422, 79.5088, 78.3107 aos AP:96.15, 90.07, 89.31 Car AP_R40@0.70, 0.70, 0.70: bbox AP:98.2938, 95.3598, 92.7532 bev AP:94.9527, 91.0876, 88.8033 3d AP:91.7655, 83.3217, 80.2945 aos AP:98.26, 95.19, 92.50 Whereas, I do inference on KITTI's testing sub-dataset, and submit it to KITTI's official website, and get the result below: Car (Detection) | 95.92 % | 94.59 % | 90.12 % Car (Orientation) | 95.90 % | 94.45 % | 89.88 % Car (3D Detection) | 87.14 % | 78.47 % | 73.53 % Car (Bird's Eye View) | 92.21 % | 88.71 % | 83.72 % Pedestrian (Detection) | 62.09 % | 52.25 % | 49.75 % Pedestrian (Orientation) | 58.38 % | 48.11 % | 45.50 % Pedestrian (3D Detection) | 45.75 % | 37.78 % | 35.39 % Pedestrian (Bird's Eye View) | 50.23 % | 42.84 % | 40.55 % Cyclist (Detection) | 85.42 % | 71.23 % | 66.10 % Cyclist (Orientation) | 85.04 % | 70.70 % | 65.52 % Cyclist (3D Detection) | 76.24 % | 60.51 % | 54.57 % Cyclist (Bird's Eye View) | 80.50 % | 65.60 % | 59.44 % Any misstake I done?

carrolltm commented 1 year ago

I met the same problem, did you solve it?

liyognxu commented 1 year ago

I git the code, and train it with 1 3090 GPU, don't even get good training results. Car AP@0.70, 0.70, 0.70: bbox AP:90.7990, 90.1060, 89.3615 bev AP:90.3149, 88.5166, 86.9862 3d AP:89.0253, 79.1707, 78.0748 aos AP:90.79, 90.01, 89.20 Car AP_R40@0.70, 0.70, 0.70: bbox AP:96.3080, 94.8953, 92.6039 bev AP:93.2292, 89.5445, 88.7421 3d AP:89.5771, 82.7497, 79.9741 aos AP:96.28, 94.77, 92.42

lubinBoooos commented 1 year ago

thks, I will try it.

mikasa3lili @.***> 于2023年6月29日周四 12:27写道：

May be you should train each class seperately, then you could get higher results. I train the model multi-class together and also only get 78+% on Mod.

— Reply to this email directly, view it on GitHub https://github.com/yifanzhang713/IA-SSD/issues/54#issuecomment-1612413055, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGTXPWX7I3DJF5LYEG5N4TTXNT74HANCNFSM6AAAAAAXB7YTJY . You are receiving this because you authored the thread.Message ID: @.***>