dddzg/up-detr
PyTorch implementation of UP-DETR for unsupervised pre-training of detection transformers on COCO object detection.
Not currently ranked — collecting fresh signals.
star history
UP-DETR introduces random query patch detection as a pretext task to pre-train transformers for object detection without human annotations. The approach inherits the DETR architecture with ResNet-50 backbone and transformer encoder-decoder, achieving 43.1 AP on COCO after fine-tuning. Pre-training can initialize CNN weights from self-supervised methods like SwAV while freezing them during transformer pre-training.
Frequently asked
- What is dddzg/up-detr?
- PyTorch implementation of UP-DETR for unsupervised pre-training of detection transformers on COCO object detection.
- Is up-detr open source?
- Yes — dddzg/up-detr is open source, released under the Apache-2.0 license.
- What language is up-detr written in?
- dddzg/up-detr is primarily written in Python.
- How popular is up-detr?
- dddzg/up-detr has 489 stars on GitHub.
- Where can I find up-detr?
- dddzg/up-detr is on GitHub at https://github.com/dddzg/up-detr.