hrnet / hrnet-bottom-up-pose-estimation Goto Github PK

Hi,

When I check the GFLOPs with the model, I found that the model with pose_hrnet_w32 backbone and 512 512 input size only have 45.03 GFLOPs.

But it showed 63.7 GFLOPs in your repo, so I am a little confused with it.

If I was wrong, please tell me how to fix it.

Thanks

Glad to see your work, it's so great : )
But when i try to use multi-scale to test the model, it comes two problems:

1. the name of param is not the same as REDEME says, may be the SACLE_FACTOR;
2. i try to change value of SCALE_FACTOR, if scale num greater than 1, there would be an error that the dim is mismatch.
   So i guess that the multi-scale test is not complete, or i didn't find the proper way to use it.
   Looking forward for your reply, Thanks : )

I have a dataset which is with skeleton annotations.
However, there are multi-person. Not single person.
How can I train this?

I do not want to first detect the bounding box of the person.

            area[i, 0] = obj['bbox'][2]*obj['bbox'][3]

This method requires bbox during training. What if the training pairs are without bbox?

How can we calculate the area and loss of offset?

Hi！
I doubt the train.py:
line 258:
perf_indicator = epoch

but in the line 251:
for epoch in range(begin_epoch, cfg.TRAIN.END_EPOCH)

Do you mean that after training, no test is performed to select the best model and the latest model defaults as the best model?

Hi.

First of all thank you very much for your work, great pose estimation libary!

When I would like to test your crowdpose trained models, I get the following error:

`root@lsdl030x:/workspace# python tools/inference_demo.py --cfg experiments/inference_demo.yaml --videoFile crowd_issue_example.mp4 --outputDir output --visthre 0.3 TEST.MODEL_FILE model/pose_crowdpose/pose_hrnet_w32_reg_delaysep_bg01_stn_512_adam_lr1e-3_crowdpose_x300.pth

=> loading model from model/pose_crowdpose/pose_hrnet_w32_reg_delaysep_bg01_stn_512_adam_lr1e-3_crowdpose_x300.pth
Traceback (most recent call last):

File "tools/inference_demo.py", line 267, in
main()
File "tools/inference_demo.py", line 190, in main
cfg.TEST.MODEL_FILE), strict=False)
File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 777, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for PoseHigherResolutionNet:
size mismatch for transition_reg.0.weight: copying a param with shape torch.Size([256, 494, 1, 1]) from checkpoint, the shape in current model is torch.Size([256, 497, 1, 1]).
size mismatch for final_layers.0.weight: copying a param with shape torch.Size([14, 32, 1, 1]) from checkpoint, the shape in current model is torch.Size([17, 32, 1, 1]).
size mismatch for final_layers.0.bias: copying a param with shape torch.Size([14]) from checkpoint, the shape in current model is torch.Size([17]).
size mismatch for final_layers.1.weight: copying a param with shape torch.Size([29, 256, 1, 1]) from checkpoint, the shape in current model is torch.Size([35, 256, 1, 1]).
size mismatch for final_layers.1.bias: copying a param with shape torch.Size([29]) from checkpoint, the shape in current model is torch.Size([35]).`

It seems like PyTorch was unable to load the model weights, to the defined model, because unmatching layer dimensions. I followed your installation instructions throughly and I can test the coco trained models.

Do you have any idea why this occurs or can you guide me towards solving this issue? Thank you in advance!

Peter

I followed the readme setup steps and came across this problem.

How can I solve that? I use Ubuntu18.04 and cuda9.0.

Hi! First of all: congratulations for the amazing work!

This is not an issue per say, but a have a dataset generated from openpose that gives me all keypoints from people in my videos, but no bounding boxes.

Is there a way to train your neural network with a ['bbox'] =[] key inside my jsons?

I retrain your model, but I found the model can't get 67.3 with single scale. (ps. I only get 60.9 without the rescore model) Can you present more training detials? Thanks!

Hi all,
Thank you for your excellent work and for sharing your code and models. I want to use your model as part of my system, here is the system's description:
Input ---> Person Object Detection ---> Pose estimation for each detected BB ---> further analysis ---> output

I want to extract pose estimation for each BB but I think I will encounter two problems:

1. As most DL models, since BB is much smaller than the image sizes you trained your model on it won't perform well. (This is a common generalization problem in DL)
2. BBs may differ in size with can cause a discrepancy for the model.

I thought of one solution where I can pad each BB to match the size you trained your model on, but I thought there must be a more elegant way to approach this.
Have you tried to tackle such problem? Do you have a suggestion for me?

Cheers.

请问这个可以在windows上运行吗？

hello authors,

Sorry i want to ask a question related to Multi-person detection.
I want to ask wheter the COCO person dataset is specialized for multi-person?
i have trained a simple model using FCN for face keypoint regression. Can you please tell me how we can get heatmaps of muliple faces in the bottom-up approach. My current model takes input of 96x96x1 image and gives 96x96x15 size heatmaps for 15 keypoints. I trained my model using datset consisting of images with single face. Do i need the datset with multiple faces? and do I need bouding box information or mask information too?

Please give me your advice
thank you

1. MULTIPROCESSING_DISTRIBUTED=False
   RuntimeError: Expected one of cpu, cuda, mkldnn, opengl, opencl, ideep, hip, msnpu device type at start of device string: 3

sorry to bother you, but I don't really understand the meaning of the 'rescore' part: score data? training rescore model ?