feat: add CIGPose pose estimation models

[wep21]

What

Adds support for CIGPose, SimCC-based human pose estimation models exported to ONNX. Since the model architecture shares its I/O contract with RTMPose, CIGPose is a type alias of RTMPose with its own model configs.

Model Variants

Config	Keypoints	Input	Scale
cigpose_17_l	17 (COCO body)	256x192	L
cigpose_17_l_384	17 (COCO body)	384x288	L
cigpose_133_l	133 (COCO-WholeBody)	256x192	L
cigpose_133_l_384	133 (COCO-WholeBody)	384x288	L
cigpose_133_x_384	133 (COCO-WholeBody)	384x288	X

All variants are available in FP32 and FP16, hosted at wep21/assets cigpose release.

Changes

• src/models/vision/cigpose/: model configs (CIGPose = RTMPose alias + Config::cigpose_* constructors)
• examples/pose-estimation/cigpose.rs + main.rs: cigpose subcommand with --scale, --is-coco, --hires, --dtype, --device options
• examples/pose-estimation/README.md, docs/model-zoo/pose.md: docs

Usage

cargo run -F cuda-full --example pose-estimation -- cigpose --dtype fp16 --device cuda:0 --processor-device cuda:0

Inference results for all variants will be posted as a follow-up comment.

🤖 Generated with Claude Code

[wep21]

Inference Results (all variants, CUDA)

Environment: NVIDIA GeForce RTX 4070 Laptop GPU (8 GB), driver 595.71.05, CUDA EP via ort v2.0.0-rc.12
Source: ./assets/bus.jpg (YOLO v26-n detector → 4 persons → batched pose estimation)

Command

cargo run -r -F cuda-full --example pose-estimation -- cigpose \
  --scale {l,x} --is-coco {true,false} --hires {true,false} \
  --dtype {fp32,fp16} --device cuda:0 --processor-device cuda:0

Performance

Variant	Flags	Dtype	Preprocess	Inference	Postprocess
cigpose-17-l	--scale l --is-coco true --hires false	fp32	451.3µs	4.453ms	28.7µs
cigpose-17-l		fp16	364.5µs	2.717ms	30.7µs
cigpose-17-l-384	--scale l --is-coco true --hires true	fp32	455.3µs	5.064ms	39.3µs
cigpose-17-l-384		fp16	567.2µs	3.461ms	39.6µs
cigpose-133-l	--scale l --is-coco false --hires false	fp32	375.2µs	4.419ms	192.0µs
cigpose-133-l		fp16	446.2µs	3.278ms	190.2µs
cigpose-133-l-384	--scale l --is-coco false --hires true	fp32	528.3µs	5.763ms	291.9µs
cigpose-133-l-384		fp16	705.4µs	3.932ms	271.7µs
cigpose-133-x-384	--scale x --is-coco false --hires true	fp32	631.6µs	10.338ms	301.1µs
cigpose-133-x-384		fp16	685.2µs	6.002ms	317.7µs

Times are per-batch averages over the 4 detected persons (3 dry-runs before timing). fp16 gives a ~1.5–1.7× inference speedup across all variants.

All 10 runs (5 variants × fp32/fp16) completed successfully.

Note: the fp16 ONNX assets were generated with NVIDIA TensorRT Model Optimizer's autocast (python -m modelopt.onnx.autocast --low_precision_type fp16 --keep_io_types --nodes_to_exclude "/gcn/.*"). The GCN head is kept in fp32 because ORT's CUDA BatchNorm<MLFloat16> kernel requires fp32 scale/bias/mean/var (standalone BatchNormalization nodes with fp16 params abort via a GSL contract violation), and the If subgraphs in the GCN don't survive fp16 type inference.

Annotated Outputs

Variant	fp32	fp16
cigpose-17-l
cigpose-17-l-384
cigpose-133-l
cigpose-133-l-384
cigpose-133-x-384

🤖 Generated with Claude Code

[wep21]

@jamjamjon Hi, I added a new model from https://github.com/53mins/CIGPose. (onnx models come from https://github.com/namas191297/cigpose-onnx). Could you check these model are applicable for usls?
CI failure seems to be unrelated to this change.

[jamjamjon]

@jamjamjon Hi, I added a new model from https://github.com/53mins/CIGPose. (onnx models come from https://github.com/namas191297/cigpose-onnx). Could you check these model are applicable for usls? CI failure seems to be unrelated to this change.

@wep21 This is amazing! CigPose looks like the new SOTA model. Thank you so much for the contribution and the PR — it’s now merged!