feat: support HRNet DarkPose pose estimation (w32/w48, COCO body & wholebody)

[wep21]

Summary

Adds HRNet top-down heatmap-based pose estimation, supporting COCO body (17 keypoints) and COCO-WholeBody (133 keypoints) with the w32 / w48 backbones at 256x192 and 384x288 input resolutions.

What's included

• src/models/vision/hrnet/impl.rs — HRNet model. Reuses the top-down affine preprocessing (person crop → fixed input size), and decodes the model's heatmap output [batch, nk, h, w] via per-keypoint argmax plus quarter-pixel sub-pixel refinement (shift 0.25 toward the larger neighbor), then maps heatmap coordinates back to the crop using center/scale.
• src/models/vision/hrnet/config.rs — Config::hrnet*() builders covering w32/w48 × body/wholebody × 256x192/384x288.
• src/models/vision/hrnet/mod.rs, module registration and re-export in src/models/vision/mod.rs.
• examples/pose-estimation/hrnet.rs + main.rs — hrnet subcommand with --width w32|w48, --is-coco, --hires, and --model (local file override). Uses a YOLO person detector for the top-down crops.

Verification

Built clean (cargo build --lib and cargo build --example pose-estimation) and ran end-to-end against a sample end2end.hrnet_w48 COCO-WholeBody ONNX (input [1,3,384,288] → output [1,133,96,72]) on assets/bus.jpg; rendered skeletons align accurately on all detected people.

cargo run --example pose-estimation -- --source ./assets/bus.jpg hrnet --width w48 --is-coco false --model ./end2end.hrnet_w48.onnx

🤖 Generated with Claude Code

[wep21]

HRNet (DarkPose) inference results

All weights are fetched on-demand from the wep21/assets hrnet release into the local cache (the per-model hub_owner/hub_repo override added in this PR). fp16 variants resolve automatically via --dtype fp16.

Source image: ./assets/bus.jpg. Person detection by YOLO; pose by HRNet.

COCO body (17 keypoints)

Variant	Command	Result
w32 · 256×192	cargo run --release --example pose-estimation -- hrnet --width w32 --is-coco true --hires false
w32 · 384×288	cargo run --release --example pose-estimation -- hrnet --width w32 --is-coco true --hires true
w48 · 256×192	cargo run --release --example pose-estimation -- hrnet --width w48 --is-coco true --hires false
w48 · 384×288	cargo run --release --example pose-estimation -- hrnet --width w48 --is-coco true --hires true

COCO-WholeBody (133 keypoints)

Variant	Command	Result
w32 · 256×192	cargo run --release --example pose-estimation -- hrnet --width w32 --is-coco false
w48 · 384×288	cargo run --release --example pose-estimation -- hrnet --width w48 --is-coco false

Add --dtype fp16 to any command to use the half-precision weights.

Verified on CPU (macOS). All six configs fetch, run, and annotate successfully. The fp16 variants were also confirmed (e.g. hrnet --width w48 --is-coco false --dtype fp16 resolves and runs …-dark-fp16.onnx, producing the same 133-keypoint result).

[wep21]

@jamjamjon Hi, is it applicable to add this new pose heatmap model? Currently, onnx is fetched from my assets repo.

[jamjamjon]

@jamjamjon Hi, is it applicable to add this new pose heatmap model? Currently, onnx is fetched from my assets repo.

Hi @wep21, thanks for the great PR! The HRNet model is definitely applicable and the code quality is solid. I merged it with a few adjustments:

• Dropped the hub_owner/hub_repo fields to keep ORTConfig lean — explicit GitHub URLs work just as well for third-party assets.
• Fixed a small URL bug in hrnet_w48_17_384() (was referencing w32 instead of w48).
• Extended ORTConfig::try_commit() so explicit GitHub URLs also get automatic dtype variant lookup (e.g. -fp16), which previously only worked for bare filenames.

The ONNX files stay in your wep21/assets repo — no need to move them. Everything compiles clean and CI passes. Thanks again for contributing!