XinNUS/CycleGRPO-RL-Data

CycleGRPO RL Data

Reinforcement-learning data for CycleGRPO (code): 19,655 DenseWorld region annotations on V3Det images and 1,000 gRefCOCO no-target expressions on COCO train2014 images.

A cycle sample carries no reference caption. It gives an image and its target regions as SAMTok mask tokens; the model captions a region, re-grounds its own caption, and the reward is how well the re-grounded region matches the target.

Files

data/
  denseworld_multi_4856.parquet     # 4,856 region-caption cycle samples (V3Det images)
  denseworld_multi_4920.parquet     # 4,920 region-caption cycle samples (V3Det images)
  denseworld_multi_4938.parquet     # 4,938 region-caption cycle samples (V3Det images)
  denseworld_multi_4941.parquet     # 4,941 region-caption cycle samples (V3Det images)
  gres_no_target_1000.parquet       # 1,000 "No target." queries (COCO train2014 images)
images/
  crops.tar                         # 66,492 zoom-in crops referenced by the DenseWorld parquets
image_lists/
  v3det.txt                         # 19,655 V3Det images used (not redistributed)
  coco_train2014.txt                # 605 COCO train2014 images used (not redistributed)
prepare_images.py                   # links the V3Det/COCO images into images/

Columns (all files):

Column	Content
images	image paths relative to this directory; the first is the full image, the rest are zoom-in crops
cap_problem	captioning prompt with one <image> per entry in images and the target region tokens
seg_answer	<answer> + target region tokens (<|mt_start|><|mt_AAAA|><|mt_BBBB|><|mt_end|>), or <answer>No target.</answer>
source	denseworld_multiple or gres_no_target
cap_answer, seg_problem, masks	unused, null

Each DenseWorld row is one V3Det image with all of its target regions, 4.4 to 5.0 of them on average. A region whose box covers less than 30% of the image also has a zoom-in crop, appended to images as an extra <image> and introduced by a Zoom in on ... clause in cap_problem.

Region tokens are SAMTok 256x2 codes (0 <= AAAA < 256 <= BBBB < 512); decode them with the mask_tokenizer_256x2.pth / sam2.1_hiera_large.pt weights from zhouyik/Qwen3-VL-4B-SAMTok.

Setup

V3Det and COCO images belong to their Flickr authors and may not be redistributed, so this repository lists the images used instead. Download V3Det from the official release and COCO 2014 train images from cocodataset.org, then:

tar -xf images/crops.tar
python prepare_images.py --v3det_images /path/to/V3Det/images \
    --coco_train2014 /path/to/coco/train2014

prepare_images.py symlinks (or with --copy, copies) exactly the listed images and reports any that are missing.

Training

Point the launcher at this directory and pass the five parquets:

D=/path/to/CycleGRPO-RL-Data
IMAGE_DIR=$D TRAIN_FILES=$D/data/denseworld_multi_4856.parquet,$D/data/denseworld_multi_4920.parquet,$D/data/denseworld_multi_4938.parquet,$D/data/denseworld_multi_4941.parquet,$D/data/gres_no_target_1000.parquet \
  bash projects/rl/qwen3vl_4b_mt.sh

The data loader draws one target at random from each DenseWorld row on every pass, together with that target's zoom-in crop. Verify every row first with projects/rl/datasets/check_rl_parquet.py --image_dir $D.

Provenance and licenses

• DenseWorld regions come from SAMTok's DenseWorld mask-generation data on V3Det images (annotations CC BY 4.0). The zoom-in crops in crops.tar are cut from those V3Det images; the underlying photos remain subject to the Flickr Terms of Use.
• No-target queries are built from gRefCOCO expressions on COCO train2014 images.
• The parquet annotations in this repository are released under CC BY 4.0.

Citation

@inproceedings{cyclegrpo2026,
  title     = {Actor as Its Own Critic: Unifying Region Understanding and Localization via CycleGRPO},
  author    = {Zhang, Xin and Wang, Haochen and Zhou, Yikang and Wang, Zhuochen and Li, Xiangtai and Tan, Robby T.},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}