This repository provides enhancements to the FCOS (Fully Convolutional One-Stage Object Detection) model, focusing on improved accuracy through techniques like center sampling and advanced loss functions (GIoU, Linear IoU). It targets researchers and practitioners in computer vision seeking to boost object detection performance.

How It Works

FCOS_PLUS builds upon the original FCOS architecture by incorporating "center sampling" during training, which prioritizes samples closer to the object's center. Additionally, it explores the use of linear IoU and GIoU loss functions, aiming to provide more accurate bounding box regression compared to standard IoU loss. These modifications are integrated into the existing FCOS framework, allowing for direct comparison and evaluation.

Quick Start & Requirements

• Installation instructions are detailed in INSTALL.md and are identical to the original FCOS project.
• Requires a PyTorch environment. Specific hardware requirements (e.g., GPU count) depend on the training configuration.
• Training command example: python -m torch.distributed.launch --nproc_per_node=8 --master_port=$((RANDOM + 10000)) tools/train_net.py --skip-test --config-file configs/fcos/fcos_R_50_FPN_1x_center_giou.yaml DATALOADER.NUM_WORKERS 2 OUTPUT_DIR training_dir/fcos_R_50_FPN_1x_center_giou

Highlighted Details

• Achieves up to 42.5 AP (minival) with FCOS_R_101_FPN_2x_center_giou, a notable improvement over the base FCOS.
• Supports multi-scale training for enhanced robustness.
• Offers pre-trained models for various configurations.
• Demonstrates the impact of different loss functions (Linear IoU, GIoU) on detection accuracy.

Maintenance & Community

• The project appears to be a research contribution, with no explicit mention of active maintenance or community channels.
• Citations are provided for the original FCOS paper.

Licensing & Compatibility

• Licensed under the 2-clause BSD License for academic use.
• Commercial use requires direct contact with the authors.

Limitations & Caveats

The project is primarily focused on research enhancements and may not be production-ready without further integration or testing. Commercial use is restricted without explicit permission.