NuScenes val mIoU ≈ 4.9% #170
Description
Hi, I evaluated PTv3 on nuScenes but get very low mIoU. Details below.
weights
config
dataset
- downloaded and preprocessed as described here: https://github.com/Pointcept/Pointcept?tab=readme-ov-file#nuscenes
command
python tools/test.py --config-file configs/nuscenes/semseg-pt-v3m1-0-base.py --num-gpus 2 --options weight=ckpt/nuscenes-semseg-pt-v3m1-0-base/model_best.pth
result
Val result: mIoU/mAcc/allAcc 0.0490/0.1642/0.2441
Class_0 - barrier Result: iou/accuracy 0.0507/0.0630
Class_1 - bicycle Result: iou/accuracy 0.0073/0.0074
Class_2 - bus Result: iou/accuracy 0.0004/0.0004
Class_3 - car Result: iou/accuracy 0.0000/0.0000
Class_4 - construction_vehicle Result: iou/accuracy 0.0000/0.0000
Class_5 - motorcycle Result: iou/accuracy 0.0028/0.0028
Class_6 - pedestrian Result: iou/accuracy 0.0241/0.0241
Class_7 - traffic_cone Result: iou/accuracy 0.0107/0.5553
Class_8 - trailer Result: iou/accuracy 0.0044/0.0185
Class_9 - truck Result: iou/accuracy 0.0415/0.0530
Class_10 - driveable_surface Result: iou/accuracy 0.0256/0.0257
Class_11 - other_flat Result: iou/accuracy 0.0000/0.0000
Class_12 - sidewalk Result: iou/accuracy 0.1332/0.8223
Class_13 - terrain Result: iou/accuracy 0.0110/0.0113
Class_14 - manmade Result: iou/accuracy 0.1435/0.1511
Class_15 - vegetation Result: iou/accuracy 0.3282/0.8927
question
I don’t understand why the accuracy is so low.
If there are any additional steps or processing required during testing, please let me know.
Thank you!