search

yancie-yjr / DBQ-SSD

The official implementation of the paper DBQ-SSD: Dynamic Ball Query for Efficient 3D Object Detection (ICLR 2023)
Apache License 2.0
17 stars 3 forks source link

fail to reproduct the performance shown in paper #3

Closed OuyangJunyuan closed 11 months ago

OuyangJunyuan commented 1 year ago

Could you give any advices to me? I think this kind of gate-based dynamic networks seem hard to converge.

Car AP@0.70, 0.70, 0.70:
bbox AP:66.2887, 50.9406, 50.1121
bev  AP:54.5814, 46.2114, 40.4017
3d   AP:47.3604, 36.0716, 31.0154
aos  AP:65.17, 49.90, 48.77
Car AP_R40@0.70, 0.70, 0.70:
bbox AP:67.3332, 51.9216, 47.9794
bev  AP:55.7742, 43.6306, 38.8716
3d   AP:44.8888, 33.6227, 29.7601
aos  AP:66.07, 50.69, 46.51
Car AP@0.70, 0.50, 0.50:
bbox AP:66.2887, 50.9406, 50.1121
bev  AP:68.2305, 58.6677, 51.9742
3d   AP:67.9217, 57.8087, 51.2324
aos  AP:65.17, 49.90, 48.77
Car AP_R40@0.70, 0.50, 0.50:
bbox AP:67.3332, 51.9216, 47.9794
bev  AP:70.9956, 55.7373, 51.5859
3d   AP:70.3942, 55.2152, 50.9204
aos  AP:66.07, 50.69, 46.51
Pedestrian AP@0.50, 0.50, 0.50:
bbox AP:33.6640, 29.2730, 27.3461
bev  AP:27.9636, 24.0015, 22.7002
3d   AP:24.5179, 21.3948, 20.0227
aos  AP:20.91, 18.85, 17.75
Pedestrian AP_R40@0.50, 0.50, 0.50:
bbox AP:30.7358, 25.6456, 23.6849
bev  AP:24.0143, 19.3956, 18.1442
3d   AP:20.2534, 16.4522, 14.7456
aos  AP:15.12, 12.55, 11.53
Pedestrian AP@0.50, 0.25, 0.25:
bbox AP:33.6640, 29.2730, 27.3461
bev  AP:39.6583, 34.1208, 32.0018
3d   AP:39.6332, 34.0367, 31.9339
aos  AP:20.91, 18.85, 17.75
Pedestrian AP_R40@0.50, 0.25, 0.25:
bbox AP:30.7358, 25.6456, 23.6849
bev  AP:37.2498, 31.2550, 28.8028
3d   AP:37.2306, 31.1573, 28.7183
aos  AP:15.12, 12.55, 11.53
Cyclist AP@0.50, 0.50, 0.50:
bbox AP:46.2147, 37.0968, 35.8150
bev  AP:41.0650, 32.3202, 31.0648
3d   AP:39.3470, 31.4047, 29.8415
aos  AP:34.72, 27.52, 26.69
Cyclist AP_R40@0.50, 0.50, 0.50:
bbox AP:44.9303, 34.4198, 33.3809
bev  AP:39.6938, 29.3634, 27.9016
3d   AP:37.9765, 28.0044, 26.6262
aos  AP:31.65, 22.93, 22.31
Cyclist AP@0.50, 0.25, 0.25:
bbox AP:46.2147, 37.0968, 35.8150
bev  AP:47.3290, 37.4161, 35.6393
3d   AP:47.2946, 37.4113, 35.6393
aos  AP:34.72, 27.52, 26.69
Cyclist AP_R40@0.50, 0.25, 0.25:
bbox AP:44.9303, 34.4198, 33.3809
bev  AP:46.0618, 34.6406, 33.3011
3d   AP:45.9248, 34.6362, 33.2539
aos  AP:31.65, 22.93, 22.31
yancie-yjr commented 1 year ago

Can your pytorch version follow us recommends?

OuyangJunyuan commented 1 year ago

Can your pytorch version follow us recommends?

thanks for your advice, I will try it again.

yancie-yjr commented 11 months ago

I'm sorry for not getting back to you sooner. In the path "DBQ-SSD/tree/main/tools/cfgs/kitti_models/DBQ-SSD.yam", the dynamic parameter of "scale" should be set to 0.1, which controls the trade-off between accuracy and latency.