Robust Multi-Human Pose Estimation

I’m looking for a developing a research oriented model that can accurately recover every joint in crowded, static images where people overlap, hide behind objects, or appear only partially. The emphasis is on distribution-robust performance under severe occlusion. I am mainly looking for large models to be used e.g., Mamba / hybrid Transformer models for human pose estimation.

The approach I have in mind mixes large Transformer backbones with geometric priors think part-affinity refinements, kinematic graph constraints, or similar. Frameworks such as PyTorch, TensorFlow, Detectron2, or MMPose are all fine as long as the pipeline stays fully reproducible on a single GPU.

Deliverables • An occlusion-aware pose estimation model with source code and training scripts • Pre-trained weights plus a clear read-me on how to reproduce results end-to-end • A concise tech report detailing architecture choices, training schedule, metrics, and ablation • A full article drafted and structured result and comparison.

Acceptance criteria

1. At least a acceptance % AP boost over my current baseline on a hidden test split containing 30 % occluded joints, crowedPose dataset can be used.
2. Inference speed ≥ 8 FPS on an RTX 3090 (batch size = 1)
3. 100 % open-source dependencies, no paid libraries or closed models

Any open source datasets, such as CrowedPose, COCO etc can be used. Budget: USD 250–750 Skills: Machine Learning (ML), Software Development, Image Processing, Open Source, Research and Development, Computer Vision, Deep Learning, Technical Documentation