Please provide the following information when requesting support.
• NVIDIA RTX 6000 Ada
• BodyPoseNet Model
• trainable_v1.0
bpnet_train_m1_coco.txt (2.9 KB)
We want to know and confirmation from you that can we train the BodyPosenet Model present in nvcr.io/nvidia/tao/tao-toolkit:5.0.0-tf1.15.5 version by freezing the backbone layers of VGG architecture or not?, because currently with the spec file, we are able to finetune on whole network, as our dataset is less, we want to train on few layers.
Freezing the VGG backbone layers is not supported by the standard BodyPoseNet spec in TAO Toolkit 5.0. For a limited dataset, please fine-tune the pretrained model with a lower learning rate.
Hi Morganh,
We are having these questions, could you please help with these,
-
whether this is BodyPoseNet model is deprecated or not, if not when it is last updated ?.
-
The version that I am using is TAO 5.2(present in this container nvcr.io/nvidia/tao/tao-toolkit:5.0.0-tf1.15.5) , is there any latest TAO having the BodyPoseNet model.
-
As this BodyPoseNet only support Vgg, when i am comparing the BodyPoseNet output with the latest state of the art keypoint models like yolo26, Higher HRnet, RTM Pose, the keypoint acuuracy of the bodyposenet is very less compared to these models,
-
I have 41 unique persons with four positions, after augmentation its 5000 dataset, i am finetuning all these models and comparing the keypoint accuracy, where i find the bodyposenet model is not upto mark
Best Regards,
Venkata Sai Teja
BodyPoseNet is based on the legacy TensorFlow 1.x TAO workflow, which has been deprecated and is no longer actively maintained in the latest TAO Toolkit releases.
It is available only through the legacy TAO TensorFlow 1 container, nvcr.io/nvidia/tao/tao-toolkit:5.0.0-tf1.15.5, for existing workflows. It is not included in the latest TAO Toolkit releases.
BodyPoseNet uses an older VGG-based architecture and does not provide the modern backbone options or pose-estimation designs available in newer pose-estimation frameworks.
Hi Morganh,
Thank you for the response, one more thing we want know from your side that
- Do you have any option other than BodyPoseNet model for the Keypoint Detection, which is supported and actively maintained by NVIDIA, and if it is available, will that be able to compete with the latest state of art of the Keypoint models.
Best Regards,
Venkata Sai Teja
For 2D full-body human keypoint detection, TAO Toolkit does not currently provide another actively maintained model that is a direct replacement for BodyPoseNet.
Other NVIDIA models are intended for different tasks. For example, BodyPose3DNet is for 3D human pose estimation, FPENet is for facial landmark estimation, and PoseClassificationNet classifies actions or poses from an existing skeleton sequence. These models cannot replace a 2D full-body keypoint detector.
You may continue to fine-tune BodyPoseNet in the legacy TAO TensorFlow 1 container, nvcr.io/nvidia/tao/tao-toolkit:5.0.0-tf1.15.5, using the pretrained BodyPoseNet weights. This should be considered a legacy maintenance option for maintaining or adapting an existing BodyPoseNet workflow, rather than a recommendation for a new long-term project.
NVIDIA also provides the open-source TRT Pose project for TensorRT-accelerated real-time pose estimation on Jetson. However, it should be treated as a reference or experimental project rather than a currently supported TAO replacement or a model guaranteed to be competitive with the latest state-of-the-art pose models.
If the primary objective is the highest 2D keypoint accuracy, we recommend selecting the best-performing modern pose model from your own benchmark, such as RTMPose, HigherHRNet, or a recent YOLO pose model, while considering deployment, latency, and maintenance requirements.
Hi Morganh,
Thank you for the reply, one more thing we want to get confirmation from you that, if any other outside model can be trained or customize by using the TAO tool Framework, suppose, if i want to use YOLO or RTMPose, HigherHRNet model via TAO, will that support it or is there any other way to do this ?, can you confirm this thing to us.
Best Regards
Venkata Sai Teja
TAO Bodyposenet does not officially support YOLO or RTMPose, HigherHRNet.
But TAO source code is available, tao_tensorflow1_backend/nvidia_tao_tf1/cv/bpnet at main · NVIDIA/tao_tensorflow1_backend · GitHub. So it is technically possible to fork the codebase and implement a custom pose-estimation architecture.
However, adding a modern architecture such as HRNet, RTMPose, or a YOLO pose model is not a simple configuration or backbone change. It would require customer-owned development across the model implementation, training pipeline, data and label processing, loss functions, evaluation, inference post-processing, export, TensorRT optimization, and deployment integration.
There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks.