Object detection without the anchor box headache

What it does FCOS performs object detection using a fully convolutional one-stage architecture that treats detection as pixel-wise prediction, similar to semantic segmentation. It removes anchor boxes and their associated hyperparameters entirely. The repository provides a PyTorch implementation built on maskrcnn-benchmark, along with pre-trained models for ResNet, ResNeXt, and MobileNet backbones.

The interesting bit The paper showed that eliminating anchors did not mean sacrificing accuracy. With a ResNet-50 backbone, FCOS achieved 38.7 AP on COCO minival against Faster R-CNN’s 36.8 AP, while training in 6.5 hours versus 8.8 hours and cutting 12 ms per image at inference. The best configuration hits 49.0% AP on COCO test-dev when paired with ResNeXt-64x4d-101 and deformable convolutions.

Key highlights

• Totally anchor-free: no anchor box computation or tuning
• Outperforms Faster R-CNN (38.7 vs 36.8 AP) with ResNet-50-FPN on COCO minival
• Trains in 6.5h vs 8.8h and infers in 44ms vs 56ms per image compared to Faster R-CNN on identical hardware
• Best model reaches 49.0% AP on COCO test-dev using ResNeXt-64x4d-101 with deformable convolutions and multi-scale testing
• A real-time variant achieving 46 FPS and 40.3 AP is maintained in the separate AdelaiDet project
• ONNX export supported; also integrated into mmdetection and other third-party backbones

Caveats

• The flagship 49.0% AP result requires multi-scale testing and heavy backbone/deformable convolution machinery
• The primary code is built on maskrcnn-benchmark; the Detectron2-based implementation lives in AdelaiDet
• Training benchmarks assume 8 V100 GPUs, though 4× 1080Ti suffice for standard ResNet-50-FPN models

Verdict A solid reference implementation for anyone studying anchor-free detection or reproducing ICCV 2019 baselines. If you want a Detectron2-based build, the README points to AdelaiDet.