cfcamo/cfcamo-rl-full

CFCamo-4B (Full FT)

Qwen3-VL-4B-Instruct fine-tuned with Counterfactual Sequence Policy Optimization (CSPO) and a Counterfactual Paired Reward (CPR) to detect camouflaged objects on natural images and abstain on target-absent counterfactuals. This is the paper-main full fine-tuning checkpoint (step 126, ε=0.5 on the 4040-pair RL set).

• 📄 Paper: https://arxiv.org/abs/2606.11231
• Code & training recipe: https://github.com/suhang2000/CFCamo
• Benchmark (CF-COD) + training data: https://huggingface.co/datasets/cfcamo/CF-COD

Use it

from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image

ckpt = "cfcamo/cfcamo-rl-full"  # or a local path
processor = AutoProcessor.from_pretrained(ckpt)
model = AutoModelForImageTextToText.from_pretrained(
    ckpt, torch_dtype="auto", device_map="auto",
).eval()

SYS = (
  "You are a camouflaged object detector. Output in this exact format:\n\n"
  "<think>your reasoning here</think>\n"
  "followed by ONE of:\n"
  "  - <bbox>[x1,y1,x2,y2]</bbox>  for a single camouflaged object\n"
  "  - <bbox>[[x1,y1,x2,y2],[x3,y3,x4,y4]]</bbox>  for multiple objects\n"
  "  - <no_camouflage/>  if no camouflaged object is present\n\n"
  "Coordinates are normalized to [0, 1000] where 1000 = full image dimension."
)
USR = (
  "Identify and locate any camouflaged object in the image.\n\n"
  "In <think></think>, briefly consider scene textures, visual anomalies, "
  "and if any object blends in. Then output ONE of:\n"
  "- <bbox>[x1,y1,x2,y2]</bbox> for one object, or [[x1,y1,x2,y2],...] for multiple\n"
  "- <no_camouflage/> if no camouflaged object"
)

image = Image.open("path/to/image.jpg").convert("RGB")
messages = [
    {"role": "system", "content": SYS},
    {"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": USR},
    ]},
]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
).to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:],
                             skip_special_tokens=True)[0])

Output is one of:

<think>...</think><bbox>[x1,y1,x2,y2]</bbox>            # single box
<think>...</think><bbox>[[...],[...]]</bbox>             # multi-box
<think>...</think><no_camouflage/>                       # abstain

Box coordinates are in [0, 1000] normalized image space.

Reproduce paper numbers

git clone https://github.com/suhang2000/CFCamo && cd CFCamo
pip install -e ".[eval]"
huggingface-cli download --repo-type dataset cfcamo/CF-COD --local-dir data/cfcod
# (place upstream COD into data/cfcod/<source>/{Imgs,GT}/ — see dataset card)
huggingface-cli download cfcamo/cfcamo-rl-full --local-dir checkpoints/cfcamo-rl-full

python scripts/eval/eval_cfcod.py \
  --cf-manifest data/cfcod/test/cf_manifest_test.jsonl \
  --data-root data/cfcod \
  --models "CFCamo=checkpoints/cfcamo-rl-full" \
  --out-dir results/cfcod_eval

Training summary

• Base: Qwen/Qwen3-VL-4B-Instruct
• Cold-start SFT: 1000 paired rows (500 detect + 500 abstain) at lr 2e-5
• RL: CSPO/CPR on 4040 paired images, 4×A800 full fine-tuning, checkpoint at step 126 (ε=0.5)

Citation

@article{li2026cfcamo,
  title   = {{CFCamo}: A Counterfactual Detect-or-Abstain Framework for Camouflaged Object Detection},
  author  = {Li, Suhang and Yoshie, Osamu and Ieiri, Yuya},
  journal = {arXiv preprint arXiv:2606.11231},
  year    = {2026}
}