Abstract
This call targets continual adaptation of human pose models under domain and task shift, with emphasis on resource constraints, sparsity, and reliable long-horizon evaluation.
Scope
I’d like to work on pose models that can adapt when the camera, dataset, or set of keypoints changes.
I’m especially interested in methods that don’t need a full retraining run each time, and in tests that show what a model forgets along the way.
Open Collaboration Tasks
- Continual adaptation benchmarks for domain and class/keypoint increments in pose estimation.
- Memory- and compute-aware adaptation under strict update budgets.
- Sparsity, pruning, and compact adaptation modules for sustainable deployment.
- Evaluation of forgetting, transfer, and stability across multi-step adaptation sequences.
What Is Already Available
- PoseAdapt benchmark structure and initial method baselines.
- Existing pose modeling pipelines that can serve as adaptation backbones.
- Reproducibility-oriented training/evaluation workflow from earlier projects.
Related Papers and Resources
Collaboration Format
- Joint experimental design, protocol definition, and baselining.
- Shared implementation with transparent reporting of constraints and tradeoffs.
- Co-authored papers and open benchmark artifacts when results mature.
To start a discussion, send a short collaboration note to mukh07@dfki.de including your relevant background and the adaptation setting you want to target.