Will McNally

Evaluated using the Detectron2 repo: - Faster R-CNN with ResNeXt-101 FPN backbone gets 56.6 AP for the person category on COCO val2017. - Faster R-CNN with ResNet-101 FPN backbone gets...

Here is the speed on my TITAN Xp GPU (evaluated over the 5000 COCO validation images using a batch size of 1, including NMS, excluding rescoring network and heatmap matching):...

@jinfagang check out KAPAO, it’s pretty fast.

It does not, sorry!

This is a known issue and I do not know the cause. I did not experience it when training on TPU.

To be clear, you get this error when you run eval.py, and not generate_splits.py, correct? Did you download the preprocessed video clips? Please see info in the train instructions. Thanks...

Yes training speed is slow due to large input size. 20 minutes per epoch doesn't sound out of the ordinary but training speed depends what hardware, batch size, and model...

There are many candidate keypoint objects. We keep the one with the highest confidence. If the confidence is higher than one that was already matched, we overwrite it.

Table 1 does not use any test time augmentation (no horizontal flipping). The original papers use horizontal flipping.

Yes!