search

tusen-ai / SST

Code for a series of work in LiDAR perception, including SST (CVPR 22), FSD (NeurIPS 22), FSD++ (TPAMI 23), FSDv2, and CTRL (ICCV 23, oral).
Apache License 2.0
801 stars 102 forks source link

error in performance of ctrl #151

Closed yrt1993 closed 1 year ago

yrt1993 commented 1 year ago

hi，I tried to reproduce the results of ctrl，and I got the following results:

OBJECT_TYPE_TYPE_VEHICLE_LEVEL_1: [mAP 0.874662] [mAPH 0.867681] OBJECT_TYPE_TYPE_VEHICLE_LEVEL_2: [mAP 0.810487] [mAPH 0.803212] OBJECT_TYPE_TYPE_PEDESTRIAN_LEVEL_1: [mAP 0.885251] [mAPH 0.835387] OBJECT_TYPE_TYPE_PEDESTRIAN_LEVEL_2: [mAP 0.824893] [mAPH 0.774855] OBJECT_TYPE_TYPE_CYCLIST_LEVEL_1: [mAP 0.817697] [mAPH 0.810792] OBJECT_TYPE_TYPE_CYCLIST_LEVEL_2: [mAP 0.797658] [mAPH 0.790865]

It don't use the optional step2 and 5 to extend and remove empty. It seems that the results of vehicle and pedestrian is OK, but the metric of cyclist is much lower that those in the paper.

Can you give me some advice and your train log if possible

loss

Abyssaledge commented 1 year ago

Sorry, I cannot provide you with the training log, which is not kept. Which base bin file of the cyclist did you use?