search

NVIDIA-AI-IOT / trt_pose

Real-time pose estimation accelerated with NVIDIA TensorRT
MIT License
974 stars 291 forks source link

what is the 'person_keypoints_val2017_modified.json'? #67

Open zzwolf opened 4 years ago

zzwolf commented 4 years ago

hello, guys.

I am trying to validate the pre-trained model provided by the author, but when I execute tasks/human_pose/eval.ipynb, it is said that I do not have 'person_keypoints_val2017_modified.json'.

I noticed that instead of 17 keypoints, the author use 18 keypoints, perhaps the 'NECK' point was calculated automatically.

So I was wondering how to convert the official 'person_keypoints_val2017.json' to 'person_keypoints_val2017_modified.json'.

Thanks!

jaybdub commented 4 years ago

Hi zzwolf,

Thanks for reaching out!

That's correct, we automatically add the 'neck' key point as the midpoint between the shoulders.

To create the modified annotations, you can run this script

https://github.com/NVIDIA-AI-IOT/trt_pose/blob/master/tasks/human_pose/preprocess_coco_person.py

Please let me know if this helps or you run into any issues.

Best, John

zzwolf commented 4 years ago

Hi zzwolf,

Thanks for reaching out!

That's correct, we automatically add the 'neck' key point as the midpoint between the shoulders.

To create the modified annotations, you can run this script

https://github.com/NVIDIA-AI-IOT/trt_pose/blob/master/tasks/human_pose/preprocess_coco_person.py

Please let me know if this helps or you run into any issues.

Best, John

Great! I successfully convert 'person_keypoints_val2017.json' to 'person_keypoints_val2017_modified.json'. Thanks a lot.

However when I ran this demo on jetson nano, I only got 10FPS by using resnet18_baseline_att_224x224_A_trt.pth. Do you know any reason why so and any particular place I need to pay attention?

Best wishes zzwolf