Detic
Open-vocabulary detection trained on 21,000 ImageNet classes, no box labels needed for the new classes.
About
Detic (Detecting Twenty-thousand Classes) is Meta FAIR's elegant trick for open-vocabulary detection: train the detector's classifier on image classification data instead of box annotations, so the vocabulary grows to tens of thousands of concepts almost for free. It uses CLIP text embeddings as the classifier, which means at inference you can detect any class just by naming it. No complex box-label assignment schemes, and it works across detector architectures (CenterNet2, Deformable DETR).
The authors report a 2.4 mAP gain overall and 8.3 mAP on novel classes on the open-vocabulary LVIS benchmark, and 41.7 mAP on standard LVIS with a Swin-B backbone, closing the gap on rare categories. It also generalizes to new datasets without fine-tuning. The code is Apache-2.0, though some bundled components carry their own licenses, which the repo documents. Detic was one of the first practical zero-shot detectors and remains a common baseline.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (ECCV 2022): +2.4 mAP overall / +8.3 mAP novel classes on open-vocab LVIS; 41.7 mAP standard LVIS (SwinB)
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 tier, metrics-v1). Evidence: ECCV 2022 paper (arXiv:2201.02605) with open-vocab LVIS numbers confirmed via the Meta Research publication page and the official repo README model zoo; Apache-2.0 confirmed via the repo license section. Strengths: simple and architecture-agnostic open-vocab recipe, first detector trained on all 21k ImageNet classes, strong cross-dataset transfer. Docked for: (1) bundled components (SWIN, CLIP, LVIS API) carry mixed licenses to track; (2) zero-shot quality lags newer contrastive detectors on rare classes. License: Apache-2.0 (majority; bundled parts under MIT/custom per repo).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ECCV 2022 (arXiv 2201.02605, Jan 2022)
- · +2.4 mAP overall / +8.3 mAP novel classes on open-vocab LVIS; 41.7 mAP standard LVIS (SwinB), author-reported
- · First detector trained on all 21k ImageNet classes; works with CenterNet2 and Deformable DETR