/Tapir-Pytorch-Inference

Minimal code for Tapir model inference in Pytorch

Primary LanguagePythonApache License 2.0Apache-2.0

Tapir Pytorch Inference

tapir_inference_pytorch.mp4

Important

This is a strip down version of the original Tapir repository focused on inference.

  • Removed the JAX dependencies and the training code.
  • Make it easy to run in real-time even with a camera feed.
  • Converted tensors from 5D to 4D (only use one frame)
  • ⚠️⚠️⚠️ONNX Inference is very slow⚠️⚠️⚠️

Installation

git clone https://github.com/ibaiGorordo/Tapir-Pytorch-Inference.git
cd Tapir-Pytorch-Inference
pip install -r requirements.txt

License

The License of the original model is Apache 2.0: License

ONNX Export

⚠️⚠️⚠️ONNX Inference is very slow⚠️⚠️⚠️

python onnx_export.py

Arguments:

  • --model: Path to the model weights
  • --resolution: Input resolution (default: 640)
  • --num_points: Number of points (default: 1000)
  • --dynamic: Export with dynamic number of points (default: False)
  • --num_iters: Number of iterations, use 0 for faster inference, 4 for better results (default: 4)
  • --output_dir: Output directory (default: ./)

Examples

Video inference:

python example_video_tracking.py

Webcam inference:

python example_webcam_tracking.py

References: