xvyv99/aggpose-lite

AggPose Lite

AggPose Lite provides deployment-ready model artifacts for AggPose-L 256x192 COCO-style top-down human pose estimation.

This repository contains:

• a cleaned safetensors checkpoint converted from the original AggPose-L COCO model;
• a depth-pruned D24 student checkpoint distilled from the D40 teacher;
• a static-batch OpenVINO INT8 artifact optimized for CPU inference;
• model files intended to be used together with the xvyv99/aggpose-lite codebase. The project is based on the upstream PediaMedAI/AggPose implementation.

Model Files

File	Variant	Format	Intended Use
AggPose-L_256x192_COCO2017.safetensors	D40	PyTorch / safetensors	Full AggPose-L COCO checkpoint
AggPose-L-D24-train32k-distilled.safetensors	D24	PyTorch / safetensors	Depth-pruned distilled student
openvino/AggPose-L-D24-train32k-distilled.static-b4-map-mlp.int8.mixed.cpu.xml	D24	OpenVINO IR	Recommended CPU deployment model
openvino/AggPose-L-D24-train32k-distilled.static-b4-map-mlp.int8.mixed.cpu.bin	D24	OpenVINO weights	Weight file required by the .xml model

For CPU-oriented offline batch inference, the recommended artifact is:

openvino/AggPose-L-D24-train32k-distilled.static-b4-map-mlp.int8.mixed.cpu.xml

The corresponding .bin file must stay in the same directory.

License

This model repository is licensed under the AGPL-3.0 License, consistent with the upstream PediaMedAI/AggPose project.