🚀 Finers

Code for paper "FineRS: Fine-grained Reasoning and Segmentation of Small Objects with Reinforcement Learning" Neurips2025. ArXiv

📊 Benchmark

👉 🔥 Our Benchmark on HuggingFace

---

🛠️Framework

---

📦 Installation

# Create environment
conda create -n finers python=3.10
conda activate finers

# Project requirements
pip install -r requirements.txt

# Install PyTorch (CUDA 11.8)
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1     --index-url https://download.pytorch.org/whl/cu118

# xFormers
pip install -U xformers==0.0.29     --index-url https://download.pytorch.org/whl/cu118

# Core dependencies
pip install bitsandbytes accelerate loguru pycocotools matplotlib sam2
pip install flash-attn --no-build-isolation   # may take long, or download from GitHub releases

# Editable install
pip install -e .

---

🤖 Download Model

apt install git-lfs
mkdir ckpts && cd ckpts

git lfs install
git clone https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct

---

▶️ Run

1️⃣ LR Data Processing

python data_process/data_converter_fixed_512_gt_crop_random_region.py

---

2️⃣ LR Training

bash training_scripts/final_lr_training.sh

---

3️⃣ HR Data Processing (Two Methods)

3.1 Faster Method

python data_process/data_converter_fixed_1920_qa.py

3.2 Paper Method (Search-Based)

# Step 1: region search based on LR model
bash data_process/data_convert_1920_with_best_region_by_LR_model.sh

# Step 2: HR conversion
python data_process/data_converter_fixed_1920_qa_with_best_region.py

---

4️⃣ HR Training

bash training_scripts/final_hr_training.sh 

# or for 3.1 faster data processing
bash training_scripts/final_hr_training_faster.sh 

---

🔄 Model Conversion (HuggingFace Format)

python3 training_scripts/model_merger.py     --local_dir workdir/xxx/global_step_xxx/actor

---

🧪 Evaluation

bash eval.sh

🔥 Our Pretrained Models for Inference

https://huggingface.co/mycfhs/FineRS/tree/main

---

Acknowledgement

• Our repo is built on Seg-Zero, EasyR1 and veRL. We thank the authors for sharing their codes.
• This work utilizes models from Qwen2.5-VL and SAM2.