H19012

Can the train Script and Data for hand_pose_resnet18_att_244_244.pth be provided? I need to do transfer learning on that model.

For the pretrained W48 inside pose_coco, what dataset was used? Was a different dataset used for the W48 inside imagenet folder? Are the training conditions used in the paper same...

Is it possible to skip resize_align_multi_scale? What does it actually do?

Here( https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation/blob/b4610aecaa5cf3de3cd69bfb13c7c79c8d514c7c/lib/core/group.py#L170) the Tensor was converted to Numpy array. Is this necessary? Doesn't Torch have the required APIs like mean and norm?

Ground Truth Confidence

2

In the json files why are the annotations' confidence lower than 100%? Should not the confidence of ground truth annotation be 100%?

Rather than always going from 64 to up, can we process 640x480 images as is?

1

Rather than always going from 64 to up, can we process 640x480 images as is? I want to test how well it does without having to shrink the image to...

(device1)(pipe1) camera -> udpsink -----Network---- (pipe2) udpsrc --> display (device2) (device1)(pipe3) microphone -> udpsink -----Network---- (pipe4) udpsrc --> speaker ( device2) I want the above in a way that ensures...

Evaluation is very very slow.

2

After each epoch evaluation is run but it takes many hours. Any way to speed it up?

Can this be jerry-rigged to work with heatmaps?