v0.3.2

PyTorch training code added.

We release a new NLF-L model (EfficientNetV2-L backbone), trained entirely with the new PyTorch training code, for much longer (1.6 million update steps) with an improved recipe, significantly outperforming the previous NLF-L model.

The newly released code also enables the use of DINOv2 ViT backbones in training. The resulting NLF-ViTg model will be released soon.

v0.2.2

• Fix wrong translation offset and 2D projection in detect_smpl_batched

v0.2.1

Small improvements using the same base crop model.

• New extra_boxes argument to inject additional bounding boxes in addition to the ones found by the built-in detector. The union of detected and extra boxes is non-maximum suppressed together.
• Better handling of buffers via nn.Buffer, ensuring proper behavior on model.to(device) or model.to(dtype)
• Accept also float/half dtype images, in which case pixel values should be between 0 and 1.
• Accept also uint16 dtype images, with pixel values ranging from 0 to 65535.
• Handle also Kannala-Brandt fisheye distorted images by internally undistorting them if distortion_coeffs have parameter vector length of 4. If distortion_coeffs is given but the length is not 4, the Brown-Conrady (max 12 params) model is applied.
• Expose the detector's flip augmentation arguments as detector_flip_aug and detector_both_flip_aug (the latter does both horizontal and vertical flip test time augmentation). TTA in the detector means that the flipped version and the original version both go through the detector, the flipped boxes are back-flipped to the original frame and the resulting boxes are all non-maximum suppressed together.

NLF v0.2.0 - Models for PyTorch and TensorFlow

Add code for canonical space setup

NLF initial model release as TensorFlow SavedModels

The models are released for noncommercial research use only.