End-to-end detection with a handful of learnable proposals

What it does SparseR-CNN is an end-to-end object detector built on Detectron2 and DETR. It represents potential objects as a small fixed set of learnable proposal boxes—100 or 300—and reports COCO box AP up to 46.4. The repository includes training and evaluation configs, pre-trained weights, and benchmark logs.

The interesting bit The entire approach hinges on optimizing just a few hundred proposal boxes end-to-end. That is a remarkably sparse starting point for a detector, and the results suggest it is sufficient.

Key highlights

• ResNet-50 with 300 proposals hits 45.0 box AP; ResNet-101 with 300 proposals reaches 46.4.
• Inference speed is 22–23 FPS on a single NVIDIA V100 for ResNet-50 variants.
• Training takes 19–29 hours on 8 V100s, depending on backbone depth and proposal count.
• Code builds on Detectron2 and DETR; third-party ports exist for mmdetection, cvpods, and PaddleDetection.
• Released under the MIT license.

Caveats

• The authors report roughly 0.3 box AP noise in their benchmarks.
• Pre-trained weights live on Google Drive, with Baidu Drive and GitHub Release fallbacks provided in case links break.
• The codebase expects the COCO dataset in a specific directory layout.

Verdict Useful if you want a research-grade, end-to-end detector on Detectron2 without engineering a proposal network. Skip it if you need a dataset-agnostic training pipeline or a polished production model zoo.