I'm testing 3DETR in outdoor senerio (nuScenes 3D dataset), but met the same promblem as #28 . The metrics are all zero after training 90 epochs. I visualize the gt and pred boxes as follows. Also, I've found loss_cardinality is unsually high, showing that almost all queries are predicted as objects.
Ground truth:
Predictions:
Following are logs:
Epoch [0/90]; Iter [0/3150]; Loss 57.61; LR 1.00e-06; Iter time 11.74; ETA 10:16:18; Mem 25411.98MB
Epoch [0/90]; Iter [10/3150]; Loss 58.18; LR 1.68e-05; Iter time 4.36; ETA 3:48:15; Mem 25499.76MB
Epoch [0/90]; Iter [20/3150]; Loss 57.50; LR 3.27e-05; Iter time 4.03; ETA 3:30:28; Mem 25499.76MB
Epoch [0/90]; Iter [30/3150]; Loss 55.73; LR 4.85e-05; Iter time 3.45; ETA 2:59:22; Mem 25499.76MB
.......
Epoch [89/90]; Iter [3120/3150]; Loss 17.18; LR 1.11e-06; Iter time 3.35; ETA 0:01:40; Mem 25499.76MB
Epoch [89/90]; Iter [3130/3150]; Loss 17.07; LR 1.05e-06; Iter time 2.49; ETA 0:00:49; Mem 25499.76MB
Epoch [89/90]; Iter [3140/3150]; Loss 17.37; LR 1.01e-06; Iter time 2.45; ETA 0:00:24; Mem 25499.76MB
Training Finished.
====================Final Eval Numbers.
mAP0.25, mAP0.50: 0.00, 0.00
AR0.25, AR0.50: 0.00, 0.00
IOU Thresh=0.25
car Average Precision: 0.00
truck Average Precision: 0.00
trailer Average Precision: 0.00
bus Average Precision: 0.00
construction_vehicle Average Precision: 0.00
bicycle Average Precision: 0.00
motorcycle Average Precision: 0.00
pedestrian Average Precision: 0.00
traffic_cone Average Precision: 0.00
barrier Average Precision: 0.00
other Average Precision: 0.00
car Recall: 0.00
truck Recall: 0.00
trailer Recall: 0.00
bus Recall: 0.00
construction_vehicle Recall: 0.00
bicycle Recall: 0.00
motorcycle Recall: 0.00
pedestrian Recall: 0.00
traffic_cone Recall: 0.00
barrier Recall: 0.00
other Recall: 0.00
I'm testing 3DETR in outdoor senerio (nuScenes 3D dataset), but met the same promblem as #28 . The metrics are all zero after training 90 epochs. I visualize the gt and pred boxes as follows. Also, I've found
loss_cardinality
is unsually high, showing that almost all queries are predicted as objects.Ground truth: Predictions:
Following are logs: Epoch [0/90]; Iter [0/3150]; Loss 57.61; LR 1.00e-06; Iter time 11.74; ETA 10:16:18; Mem 25411.98MB Epoch [0/90]; Iter [10/3150]; Loss 58.18; LR 1.68e-05; Iter time 4.36; ETA 3:48:15; Mem 25499.76MB Epoch [0/90]; Iter [20/3150]; Loss 57.50; LR 3.27e-05; Iter time 4.03; ETA 3:30:28; Mem 25499.76MB Epoch [0/90]; Iter [30/3150]; Loss 55.73; LR 4.85e-05; Iter time 3.45; ETA 2:59:22; Mem 25499.76MB ....... Epoch [89/90]; Iter [3120/3150]; Loss 17.18; LR 1.11e-06; Iter time 3.35; ETA 0:01:40; Mem 25499.76MB Epoch [89/90]; Iter [3130/3150]; Loss 17.07; LR 1.05e-06; Iter time 2.49; ETA 0:00:49; Mem 25499.76MB Epoch [89/90]; Iter [3140/3150]; Loss 17.37; LR 1.01e-06; Iter time 2.45; ETA 0:00:24; Mem 25499.76MB Training Finished. ====================Final Eval Numbers. mAP0.25, mAP0.50: 0.00, 0.00 AR0.25, AR0.50: 0.00, 0.00
IOU Thresh=0.25 car Average Precision: 0.00 truck Average Precision: 0.00 trailer Average Precision: 0.00 bus Average Precision: 0.00 construction_vehicle Average Precision: 0.00 bicycle Average Precision: 0.00 motorcycle Average Precision: 0.00 pedestrian Average Precision: 0.00 traffic_cone Average Precision: 0.00 barrier Average Precision: 0.00 other Average Precision: 0.00 car Recall: 0.00 truck Recall: 0.00 trailer Recall: 0.00 bus Recall: 0.00 construction_vehicle Recall: 0.00 bicycle Recall: 0.00 motorcycle Recall: 0.00 pedestrian Recall: 0.00 traffic_cone Recall: 0.00 barrier Recall: 0.00 other Recall: 0.00