Boxes elimination may be omitted

FateScript / CenterNet-better

An easy to understand and better performance version of CenterNet

Apache License 2.0

548 stars 104 forks source link

Boxes elimination may be omitted #28

Closed raymon-tian closed 4 years ago

raymon-tian commented 4 years ago

Hi, thx for your work.

Diving into the code, I notice that during inference the detector always predicts 100 (set with topK hyper-parameter) instances per image. For the whole COCO validation dataset (5000 images), 500000 instances are saved into coco_instances_results.json and instances_predictions.pth.

Thus, I doubt that some post-process steps may be omitted. If I misunderstood, I will be appreciated if you could solve my confusion. Thx.

FateScript commented 4 years ago

COCO dataset only evaluate top 100 bounding boxes, you may check other models trained on COCO for this.