why I have very low accuracy for head position?
Closed this issue · 7 comments
Hi,
I want to calculate the angle of the head like
Briefly, I use openpose and @ArashHosseini 's fork to calculate 3d position of the head.
I recorded several very simple scenarios to observe openpose and 3d-pose-baseline results. Although openpose works very nice , the head of the 3d model is almost fixed.
I realized that we are not using neck/noise joint and I think that it is the reason of this low accuracy.
May I get your feedback/comments? How can i solve it?
Thanks,
What dataset are you using, and what precision do you have?
Actually, I created a very simple dataset using a youtube video. You can see the input video, openpose results, and 3d-pose results.
Is it possible to confirm that It is normal? If not what did I do wrong? (If you need, I can share the json and poses3d results for each frame)
In order to estimate 3d pose, I used the following command:
python src/openpose_3dpose_sandbox.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise --use_sh --epochs 200 --load 4874200 --pose_estimation_json OPENPOSE_OUTPUT_DIRECTORY --write_gif --gif_fps 24
Hi @una-dinosauria,
I know that you are overloaded but I would be glad if you can just check my results and give your opinion.
I am looking forward to hearing from you..
I see. No idea. Probably the difference between open pose and our pre-trained model is largely to blame. You could dump 2d detections for H3.6M, retrain the 2d->3d model, and try again.
Thank you @una-dinosauria for your answer
When I check this demo video, I realized that the trained model has a similar issue. Maybe it can be a good future work to improve neck and head estimation.
Closing for lack of activity. Please reopen if the issue is still ongoing.