Benchmarking Detectors for the RIVA Cervical Cytology Challenge

This repository contains the official code for Team jht010312's solution to the ISBI 2026 RIVA Cervical Cytology Challenge. We fine-tuned the Co-DETR model with ViT-coco-instance weights, which achieved 1st place in Track A (Multi-class Detection) and 3rd place in Track B (Single-class Detection).

🔗 Download our fine-tuned Co-DETR weights here

🏆 Performance Highlights

Our solution secured top placements across both evaluation tracks:

• Track A (Multi-class Detection):

  • 1st Place mAP: 0.2273 (Preliminary) / 0.1937 (Final)
  • Focus: Fine-grained classification of 8 categories (NILM, ENDO, INFL, ASC-US, LSIL, HSIL, ASC-H, SCC).

• Track B (Single-class Detection):

  • 3rd Place mAP: 0.6036 (Preliminary) / 0.5906 (Final)

🚀 How to reproduce our results

To reproduce our results, you need to set up the Co-DETR environment and use our provided configuration files and inference script.

1. Setup Co-DETR

First, clone the official Co-DETR repository and install the required dependencies:

git clone https://github.com/Sense-X/Co-DETR.git
cd Co-DETR
# Follow the installation instructions in the Co-DETR README to set up the environment

2. Prepare Configuration and Scripts

Copy the configuration files and the inference script from this repository to the Co-DETR directory structure:

• Config Files: Copy co_dino_5scale_vit_large_coco_instance_isbi*.py to Co-DETR/projects/configs/co_dino_vit/
• Inference Script: Copy infer.py to Co-DETR/tools/

3. Run Inference

Use the following command to run inference on your images:

python tools/infer.py --config projects/configs/co_dino_vit/[config_file] --weights [path_to_weights] --img-dir [path_to_images]

⚙️ Methodology

Benchmark Models

We evaluated a diverse set of architectures:

• Co-DETR: Collaborative Hybrid Assignments Training (Best performing).
• D-Fine: Distribution refinement for regression.
• EVA-02: Next-generation transformer based on masked image modeling.
• Mask R-CNN & Keypoint R-CNN: Robust two-stage baselines.

Training Details

• Framework: PyTorch 2.0.1.
• Input Resolution: 1024 × 1024 pixels.
• Augmentation: Random horizontal and vertical flipping.
• Co-DETR Configuration:
  • Epochs: 12
  • Batch Size: 2
  • Optimizer: AdamW (weight decay 0.0001)
  • Learning Rate: 0.0001 (with 500-step linear warmup).
  • Pre-training: Initialized with ViT-coco-instance weights.

Inference Strategy

Predictions are generated at 1024×1024 resolution. To meet the challenge requirement of fixed 100×100 pixel bounding boxes, we apply post-processing: if a detected box is within a 5-pixel margin of the image boundary, we dynamically pad its truncated spatial axis inward to ensure the 100-pixel dimension.

🔗 Resources and Links

• Challenge Website: RIVA Cervical Cytology Challenge
• Model Weights: Download our fine-tuned Co-DETR weights
• Codebase: Built upon Sense-X/Co-DETR.