Performance on LVIS mAP is low

[WesleyHsieh0806]

We have reproduced mAP on TAO dataset, and found that the performance is quite low.

AP	AP50	AP75	track3D mAP
8.8717	14.0362	9.4615	7.7657

The numbers are generated with the block here , while track3D mAP is calculated with the tao toolkit.

Could you provide the mAP numbers you get on TAO? We need to further verify it there is any bugs here

[siyuanliii]

Hi, thanks for your question! The block you show computes the box AP on LVIS. To get a reasonable number on that, you need to change the NMS setting and test cfg to the default LVIS one which uses intra-class NMS and allows 300 bboxes output. Our current default NMS settings are class agnostic and test cfg only allows outputting a maximum of 50 bboxes per image to optimize the TETA metric, which does not allow the high overlapping predictions in the same location.

[WesleyHsieh0806]

Thanks for replying! I modified the test_cfg to the following

test_cfg=dict(
        rcnn=dict(
            score_thr=0.0001,
            nms=dict(type='nms', iou_threshold=0.5, class_agnostic=False, split_thr=100000),
            max_per_img=300)
            )

and gets a more reasonable number of AP. However, the track3D mAP on TAO does not differ a lot.

Is there something I might be missing, such as any hyperparameters in test_cfg that might influence the results on tracking? Thank you in advance, great work!

[siyuanliii]

Thanks for your question! Track mAP by default, you need to as well output lots of overlapping false positive tracks to get a reasonable score. However, the TETer is not designed to output such results. Thus, you have to modify the tracker by yourself to make sure that the tracker still outputs those false positive overlapping tracks. It is strongly recommended that you test the tracker using TETA metric which is designed better analyze the trackers and avoid such false positive overlapping tracks.