Nobut1/nobut-swimmer-pose

NoBut swimmer pose (fp16 ONNX)

Half-precision (fp16) ONNX copy of ViTPose-L finetuned on SwimXYZ by willalexm (https://huggingface.co/willalexm/vitpose-l-swimxyz), used by NoBut to find swimmers' joints in the browser with ONNX Runtime Web (WebGPU).

• Change from the original: weights converted from fp32 to fp16 (inputs and outputs kept fp32) to halve the download. On 167 real underwater frames, half of the joints land on the same pixel as the original and nine in ten within 11 px.
• Input: (N, 3, 256, 192) ImageNet-normalized RGB crop of the swimmer. Output: (N, 17, 64, 48) heatmaps, COCO-17 order.

Credits and licenses

• Model: ViTPose-L finetuned on SwimXYZ, willalexm, Apache License 2.0.
• ViTPose: Xu et al., ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation (NeurIPS 2022), Apache License 2.0.
• Training data: SwimXYZ, Fiche et al., SwimXYZ: A large-scale dataset of synthetic swimming motions and videos (2023), CC BY 4.0.

This copy is distributed under the Apache License 2.0, like the original.