ViTPose
Plain vision transformers for pose: 81.1 mAP on COCO test-dev with a ViT and two deconvolutions.
About
ViTPose asked a deliberately naive question — can a plain, non-hierarchical vision transformer do pose estimation? — and answered it by topping the COCO leaderboard. An MAE-pretrained ViT backbone plus a trivial two-deconvolution decoder reaches 81.1 mAP on COCO test-dev (ensemble), scaling cleanly from tiny to a 1-billion-parameter ViTAE-G.
The follow-up ViTPose++ (TPAMI 2023) turned it into a foundation model for generic body pose: one backbone, task-specific heads for humans, hands, faces, and even animals, reaching 92.8 AP on OCHuman and 94.3 PCKh on MPII. Everything is Apache-2.0, and Hugging Face ships the checkpoints in transformers.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: ViTPose-G 80.9 mAP and 81.1 mAP (ensemble) on COCO test-dev; ViTPose-B 75.8 AP on COCO val; ViTPose++ 92.8 AP on OCHuman, 78.3 AP on CrowdPose. No independent replication of these figures was found.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified, metrics-v1). Evidence: peer-reviewed NeurIPS 2022 paper and TPAMI 2023 extension by named researchers; Apache-2.0 code and weights; checkpoints mirrored in Hugging Face transformers. Strengths: strong provenance, genuinely open, scales impressively. Docked for: (1) accuracy figures are provider-reported with no independent replication; (2) the large variants are compute-heavy for a pose model. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · NeurIPS 2022 (Xu, Zhang, Zhang, Tao); plain MAE-pretrained ViT + two deconvolutions.
- · Provider-claimed 81.1 mAP COCO test-dev (ensemble); ViTPose++ (TPAMI 2023) generalizes to hand/face/animal.
- · Checkpoints mirrored in Hugging Face transformers; large variants are compute-heavy.