How to obtain bbox file?

[ericzw]

Hi, could you tell me how to generate bounding box file?
Or, could you provide the bounding box files you used?

[shrubb]

Hi @qiuzhongwei-USTB,

for Human3.6M, follow the instructions in here for ground truth boxes, or see this Mask R-CNN boxes;

for CMU Panoptic, the bounding boxes are not yet ready, as is the whole pipeline, sorry for that.

[ericzw]

Thanks，
could you tell me what's the difference of using undistort images and origin image?

[shrubb]

Please read about camera models, for example, here. With original images, the ground truth 3D points will project to slightly inaccurate locations in images (still may be OK, but the quality will be lower). Undistortion corrects this.

[ericzw]

Thanks for your reply! I met some problems again.

1. I found that it's very slow when I evaluate your pre-trained model(about 70 s/img, one V100 GPU).
   Could provide your train times and test times on Human3.6 dataset?

2. There is an error that when I set 'retain_every_n_frames_in_test=5 or 64'.

error information：
Experiment name: [email protected]:24:58
5it [06:07, 73.45s/it]
/root/v-helzhe/zw/h36m/data/learnable-triangulation-pytorch/mvn/datasets/human36m.py:222: RuntimeWarning: invalid value encountered in true_divide
action_scores[k] = v['total_loss'] / v['frame_count']
/root/v-helzhe/zw/h36m/data/learnable-triangulation-pytorch/mvn/datasets/human36m.py:222: RuntimeWarning: invalid value encountered in double_scalars
action_scores[k] = v['total_loss'] / v['frame_count']

[shrubb]

Not sure. Let us look into this this after the weekend.

[shrubb]

@qiuzhongwei-USTB did you close the issue because you worked around those problems? If yes, how did you solve them?

[ericzw]

@qiuzhongwei-USTB did you close the issue because you worked around those problems? If yes, how did you solve them?

Yes, v['frame_count'] could be zeros. I change the code like the follows:
for k, v in action_scores.items():
if v['frame_count'] > 0:
action_scores[k] = v['total_loss'] / v['frame_count']
else:
action_scores[k] = 0

[shrubb]

What about the long evaluation time?

[ericzw]

I didn't solve these. I can bear the evaluation time when I use 8 GPU with distributed testing.
Can you provide your evaluation time?

[karfly]

Hi @qiuzhongwei-USTB!
It's very strange that you get such big evaluation times. I've just checked our evaluation times on full Human3.6M validation dataset:

• Algebraic: 1 hour 29 minutes
• Volumetric: 2 hours 1 minute

We use single Nvidia Tesla P40 GPU for evaluation.

Could you provide full bash command, which you used for evaluation?