Closed julianleelinker closed 2 years ago
LiDAR models do not transfer well from one dataset to another one. Here is a good read from @ltriess giving an overview on the topic.
Hi @MartinHahner, thanks for your reply, I understand your point. However the performance shouldn't be so bad, according to this article Train in Germany, Test in The USA: Making 3D Object Detectors Generalize , where the KITTI pretrained pointrcnn detector still gives AP(BEV) > 20% and AP(3D) > 10%. So I am wondering if I missed something in the config file or threre is some bug in the code such that models between datasets are not compatible.
Train in Germany, Test in The USA: Making 3D Object Detectors Generalize
Yeah, I know that paper. It's great work. Note though, they report results using PointRCNN and not PV-RCNN. Even though PV-RCNN is superior on KITTI, it does not mean that it also transfers better to Waymo. Were you able to reproduce their results for PointRCNN?
Sorry I made typos, I was actually testing using PointRCNN the result is also PointRCNN result.
OK, now you have a valid point. Maybe reach out to the authors or refer to their published code on how to train on KITTI and test on Waymo.
Hi, I have ran kitti -> waymo experiment of Train in Germany, Test in The USA: Making 3D Object Detectors Generalize, I find that a setting of pointrcnn is different from OpenPCDet and paper's code. https://github.com/cxy1997/3D_adapt_auto_driving/blob/master/pointrcnn/tools/cfgs/default.yaml#L28 So, I change kitti dataset setting https://github.com/open-mmlab/OpenPCDet/blob/c9d31d393acaae34d74cb03bd6ccff9976d3d1f3/tools/cfgs/dataset_configs/kitti_dataset.yaml#L49-L50 to
used_feature_list: ['x', 'y', 'z',],
src_feature_list: ['x', 'y', 'z',],
and use paper's code to translate waymo dataset to kitti format, and get a different result, perhaps the result is more reasonable, but has a gap with paper
2021-10-03 18:58:46,535 INFO *************** EPOCH 71 EVALUATION *****************
2021-10-03 19:04:41,980 INFO *************** Performance of EPOCH 71 *****************
2021-10-03 19:04:41,980 INFO Generate label finished(sec_per_example: 0.1394 second).
2021-10-03 19:04:41,980 INFO recall_roi_0.3: 0.592279
2021-10-03 19:04:41,981 INFO recall_rcnn_0.3: 0.597099
2021-10-03 19:04:41,981 INFO recall_roi_0.5: 0.489929
2021-10-03 19:04:41,981 INFO recall_rcnn_0.5: 0.538952
2021-10-03 19:04:41,981 INFO recall_roi_0.7: 0.130628
2021-10-03 19:04:41,981 INFO recall_rcnn_0.7: 0.286089
2021-10-03 19:04:41,982 INFO Average predicted number of objects(2549 samples): 6.011
2021-10-03 19:04:46,425 INFO Car AP@0.70, 0.70, 0.70:
bbox AP:36.8763, 38.8078, 36.5180
bev AP:39.4154, 41.3404, 37.1790
3d AP:11.1114, 11.4454, 12.1022
aos AP:36.64, 38.55, 36.30
Car AP_R40@0.70, 0.70, 0.70:
bbox AP:33.8597, 35.4379, 35.6711
bev AP:37.5705, 38.7479, 36.7095
3d AP:9.0654, 10.1663, 10.1325
aos AP:33.62, 35.19, 35.37
Car AP@0.70, 0.50, 0.50:
bbox AP:36.8763, 38.8078, 36.5180
bev AP:70.8551, 70.4873, 69.7542
3d AP:65.7294, 61.2301, 60.8189
aos AP:36.64, 38.55, 36.30
Car AP_R40@0.70, 0.50, 0.50:
bbox AP:33.8597, 35.4379, 35.6711
bev AP:73.8100, 71.9384, 67.7279
3d AP:64.7294, 63.3294, 59.4869
aos AP:33.62, 35.19, 35.37
and run stat norm code between kitti and waymo, re-train pointrcnn with kitti after stat norm, the result is similar to the paper, but also has a gap with paper.
2021-10-04 17:07:55,597 INFO *************** EPOCH 80 EVALUATION *****************
2021-10-04 17:14:03,482 INFO *************** Performance of EPOCH 80 *****************
2021-10-04 17:14:03,483 INFO Generate label finished(sec_per_example: 0.1443 second).
2021-10-04 17:14:03,483 INFO recall_roi_0.3: 0.626237
2021-10-04 17:14:03,483 INFO recall_rcnn_0.3: 0.628432
2021-10-04 17:14:03,483 INFO recall_roi_0.5: 0.581303
2021-10-04 17:14:03,483 INFO recall_rcnn_0.5: 0.593398
2021-10-04 17:14:03,483 INFO recall_roi_0.7: 0.415340
2021-10-04 17:14:03,483 INFO recall_rcnn_0.7: 0.479685
2021-10-04 17:14:03,484 INFO Average predicted number of objects(2549 samples): 5.109
2021-10-04 17:14:07,974 INFO Car AP@0.70, 0.70, 0.70:
bbox AP:51.5635, 53.4044, 48.2356
bev AP:67.8604, 61.6724, 61.0304
3d AP:47.4996, 48.2412, 47.3004
aos AP:51.19, 52.99, 47.88
Car AP_R40@0.70, 0.70, 0.70:
bbox AP:50.4320, 50.7123, 47.8743
bev AP:65.9226, 64.2689, 59.9811
3d AP:47.9117, 48.7304, 43.9152
aos AP:50.05, 50.33, 47.46
Car AP@0.70, 0.50, 0.50:
bbox AP:51.5635, 53.4044, 48.2356
bev AP:71.0432, 70.7979, 62.9648
3d AP:70.0923, 70.1391, 62.5190
aos AP:51.19, 52.99, 47.88
Car AP_R40@0.70, 0.50, 0.50:
bbox AP:50.4320, 50.7123, 47.8743
bev AP:70.8459, 68.6688, 64.0247
3d AP:70.0051, 67.9982, 63.4958
aos AP:50.05, 50.33, 47.46
By the way, I check waymo dataset after tranlated to kitti foramt with kitti_object_vis , but it seems that the location of the boxes is somewhat deviated, not sure whether this will affect the results.
This issue is stale because it has been open for 30 days with no activity.
Hi @Shuntw6096 , may I ask how you change the kitti_dataset.yaml as you mentioned? When I tried to delete the 'intensity', it will appear errors when loading the pretrained model pth file. Can you help give me some hint? Thanks in advance!
This issue is stale because it has been open for 30 days with no activity.
This issue was closed because it has been inactive for 14 days since being marked as stale.
Hi, thanks for sharing your nice work. I just try testing on waymo dataset using kitti pretrained PointRCNN model and got poor result.
Change I made in waymo_dataset.yaml:
My pointrcnn.yaml is copied from kitti_models/pointrcnn.yaml with modification of vehicle mean_size:
My result:
I encountered some dependency issue so I installed waymo-open-dataset-tf-2-1-0==1.2.0, instead of waymo-open-dataset-tf-2-0 -0==1.2.0 I am not sure if that is the cause of the poor result (it seems to me not). Do you have any idea what leads to the poor result? Thank you very much.