Detecting Twenty-thousand Classes using Image-level Supervision

Detic: A Detector with image classes that can use image-level labels to easily train detectors.

Detecting Twenty-thousand Classes using Image-level Supervision,
Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krähenbühl, Ishan Misra,
arXiv technical report (arXiv 2201.02605)


  • Detects any class given class names (using CLIP).

  • We train the detector on ImageNet-21K dataset with 21K classes.

  • Cross-dataset generalization to OpenImages and Objects365 without finetuning.

  • State-of-the-art results on Open-vocabulary LVIS and Open-vocabulary COCO.

  • Works for DETR-style detectors.


See installation instructions.


Run our demo using Colab (no GPU needed): Open In Colab

We use the default detectron2 demo interface.
For example, to run our 21K model on a messy desk image (image credit David Fouhey) with the lvis vocabulary, run

mkdir models
wget -O models/Detic_LCOCOI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.pth
python --config-file configs/Detic_LCOCOI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.yaml --input desk.jpg --output out.jpg --vocabulary lvis --opts MODEL.WEIGHTS models/Detic_LCOCOI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.pth

If setup correctly, the output should look like:

The same model can run with other vocabularies (COCO, OpenImages, or Objects365), or a custom vocabulary. For example:

python --config-file configs/Detic_LCOCOI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.yaml --input desk.jpg --output out2.jpg --vocabulary custom --custom_vocabulary headphone,webcam,paper,coffe --confidence-threshold 0.3 --opts MODEL.WEIGHTS models/Detic_LCOCOI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.pth

The output should look like:

Note that headphone, paper and coffe (typo intended) are not LVIS classes. Despite the misspelled class name, our detector can produce a reasonable detection for coffe.

Benchmark evaluation and training

Please first prepare datasets, then check our MODEL ZOO to reproduce results in our paper. We highlight key results below:

  • Open-vocabulary LVIS

    mask mAP mask mAP_novel
    Box-Supervised 30.2 16.4
    Detic 32.4 24.9
  • Standard LVIS

    Detector/ Backbone mask mAP mask mAP_rare
    Box-Supervised CenterNet2-ResNet50 31.5 25.6
    Detic CenterNet2-ResNet50 33.2 29.7
    Box-Supervised CenterNet2-SwinB 40.7 35.9
    Detic CenterNet2-SwinB 41.7 41.7
    Detector/ Backbone box mAP box mAP_rare
    Box-Supervised DeformableDETR-ResNet50 31.7 21.4
    Detic DeformableDETR-ResNet50 32.5 26.2
  • Cross-dataset generalization

    Backbone Objects365 box mAP OpenImages box mAP50
    Box-Supervised SwinB 19.1 46.2
    Detic SwinB 21.4 55.2


The majority of Detic is licensed under the Apache 2.0 license, however portions of the project are available under separate license terms: SWIN-Transformer, CLIP, and TensorFlow Object Detection API are licensed under the MIT license; UniDet is licensed under the Apache 2.0 license; and the LVIS API is licensed under a custom license (” If you later add other third party code, please keep this license info updated, and please let us know if that component is licensed under something other than CC-BY-NC, MIT, or CC0

Ethical Considerations

Detic’s wide range of detection capabilities may introduce similar challenges to many other visual recognition and open-set recognition methods.
As the user can define arbitrary detection classes, class design and semantics may impact the model output.


If you find this project useful for your research, please use the following BibTeX entry.

  title={Detecting Twenty-thousand Classes using Image-level Supervision},
  author={Zhou, Xingyi and Girdhar, Rohit and Joulin, Armand and Kr{\"a}henb{\"u}hl, Philipp and Misra, Ishan},
  booktitle={arXiv preprint arXiv:2201.02605},


View Github