3d output issue

[gascione]

Hi! We're running the following code (we've done the 150 epochs) :
python src/openpose_3dpose_sandbox.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise --use_sh --epochs 150 --load 3655650 --openpose json_directory --write_gif --gif_fps 24

This is one of the output frames from Openpose:

This is the result we get from the code:

Operating system Windows, tensorflow v 1.12.0 , python v 3.6.6

Are we missing a flag? We tried with another video and we have the same issue with the legs

Thank you a lot.

[EnricoBeltramo]

Did you use branch https://github.com/ArashHosseini/3d-pose-baseline?
Same issue with same version. I used the pre-trained network (200 epochs)

[gascione]

Yes, we used that branch. We couldn’t open the pre-trained network, therefore we trained it (150 epochs)

[EnricoBeltramo]

About the pretrained model, if you are using Windows I had the same problem (#102). I resolved extracting on a Linux system.

[EnricoBeltramo]

I guess I find the issue. Looking in Openpose 1.4.0 the default body model is changed, so don't match with actual actual 3d-pose-baseline project. Using the version of Openpose 1.3.0 (https://github.com/CMU-Perceptual-Computing-Lab/openpose/releases/tag/v1.3.0) for me works well. May be remapping the joint points or set OpenPose 1.4.0 with previous body model should work as well.

[ArashHosseini]

Happy Christmas time @gascione @EnricoBeltramo,
we have now more joints(R&L foot ankle aim) in the output of openpose, can you please post the content of the json file? I may have to make some changes, masking or involving the new joints...or if its blocking you now, fast solution would be to downgrade openpose to 1.3.0, @EnricoBeltramo has already mentioned this, are the foot ankle aim jnts necessary for you @gascione ?

[EnricoBeltramo]

Hello,
Happy late Christmas time! I post the same video processed with OpenPose 1.3.0 and 1.4.0 and the result of 3d-pose:
Openpose_1_3_0_ok.zip
Openpose_1_4_0_fail.zip

[ArashHosseini]

thx the "ok" output is using coco as default "--model_pose", the failed one is using body+foot which contains more than 18 jnts, so mapping would be possible to support COCO and Body_25 model, but we can force openpose to use the COCO model for 1.4.0,

DEFINE_string(model_pose, "BODY_25", "Model to be used. E.g., COCO (18 keypoints), MPI (15 keypoints, ~10% faster), MPI_4_layers (15 keypoints, even faster but less accurate).");

@gascione add explicitly the "model_pose" flag and COCO as target model. like:

./build/examples/openpose/openpose.bin ... --model_pose COCO

there is nothing to do on 3d-pose-baseline args, maybe adding interpolation

python src/openpose_3dpose_sandbox.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise --use_sh --epochs 150 --load 3655650 --openpose json_directory --write_gif --gif_fps 24 --interpolation --multiplier 10

[ArashHosseini]

@una-dinosauria fixed with commit 0b3400c , it will check the length of the output and mapping if needed.
@gascione can you pull and try it please, you have nothing to change on the command
python src/openpose_3dpose_sandbox.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise --use_sh --epochs 150 --load 3655650 --openpose json_directory --write_gif --gif_fps 24
@EnricoBeltramo "ok" and "fail" running now through

[una-dinosauria]

@ArashHosseini Thanks for keeping this up to date. This gets asked so often that I'll add a pointer to this project in the README.

Cheers,

[ArashHosseini]

@una-dinosauria sure, i will look after related issues, i orientate myself to the tags.
thx

[gascione]

Thanks for the answers. Using the "model_pose" flag and COCO as target model in Openpose worked perfectly. However, I tried with the new code and I still experience the same issue:

Tried with the @EnricoBeltramo outputs too and this is the result:

[ArashHosseini]

@gascione right, compared it to the array without the confidence score before, fixed with 2861831 , can i ask you for a last try so we can close it.
thx

[gascione]

Hi! It's working perfectly with the new code. Thank you so much

[ArashHosseini]

@una-dinosauria, can I ask you to change the note in the main ReadMe, now there is also the possibility to use tf-pose-estimation as an alternative to openpose, thank you.
@gascione, sounds great so maybe we can close this issue.