KRMayD/COD10K_GMPO_Used_Data

COD10K GMPO Used Data

This is a research repack of the COD10K camouflaged-object (CAM) subset used in our CLIP DPO/GMPO experiments. It is not an official COD10K distribution.

The package contains the exact original images, masks, generated negative images, and portable caption CSV files used for training and segmentation evaluation. All paths in the portable CSV files are relative to this dataset root.

Data Split

Split	Contents	Count
Train	Original CAM images	3,040
Train	Gaussian negative images	3,040
Train	Full Stable Diffusion inpaint negative images	3,040
Train	Mask-only Stable Diffusion composite negative images	3,040
Train	Object, instance, and edge masks	3,040 each
Test	Original CAM images	2,026
Test	Object, instance, and edge masks	2,026 each

The exported train and test image stem sets do not overlap.

Directory Structure

• train/images/original: CAM training images referenced by the DPO caption CSVs.
• train/images/negative_gaussian_gt_object: Gaussian-noise negative images generated inside GT object masks.
• train/images/negative_sd_inpaint_full_gt_object: full Stable Diffusion inpaint outputs.
• train/images/negative_sd_inpaint_composite_gt_object: mask-only Stable Diffusion inpaint composites, with the original image outside the mask and the inpaint result inside the mask.
• train/masks/*: COD10K training object, instance, and edge masks.
• test/images/original_cam: COD10K CAM test images used for DSC/NSD evaluation.
• test/masks/*_cam: COD10K CAM test object, instance, and edge masks.
• captions: portable caption templates, manifests, and training CSV files.
• metadata/original_cod10k_metadata: metadata copied from the local COD10K source package.
• metadata/source_file_manifest.csv: source-to-export file mapping and byte sizes.
• metadata/test_cam_manifest.csv: portable test image-to-mask mapping.
• metadata/dataset_summary.json: counts and export metadata.

Main GMPO Training CSV

The final image-specific background-caption experiment used:

captions/cod10k_train_cam_dpo_1caption_diffusion_bgpos_vs_bgonly.csv

It contains one row per training image:

• filename: original CAM image.
• filename_neg: mask-only Stable Diffusion composite negative image.
• Caption: image-specific background-only caption plus a target-presence clause.
• Caption_neg: the same image-specific background-only caption without a target-presence claim.

The package also preserves the earlier five-caption Gaussian, full-inpaint, and mask-only-composite ablation CSV files. Each of those files contains five caption pairs per training image.

Derived Content

This package modifies and extends the COD10K CAM subset by:

1. Selecting and repacking the CAM train and test samples.
2. Generating Gaussian-noise negatives inside object masks.
3. Generating Stable Diffusion inpaint negatives and mask-only composites.
4. Generating common, superclass, animal-specific, and image-specific captions.
5. Rewriting file references as portable paths relative to this dataset root.

Integrity

At export verification time:

• 29,384 image and mask files decoded successfully.
• No referenced image, negative image, or mask was missing.
• No paired image/mask dimension mismatch was found.
• No empty object mask was found.
• All JSON and JSONL sidecar files parsed successfully.

License and Terms

The original COD10K metadata identifies the data license as CC BY-NC-SA 2.0, and the official COD10K repository states that the dataset is available for non-commercial purposes only. This derived package is distributed under the same CC BY-NC-SA 2.0 license.

Users must:

• Give appropriate credit to the original COD10K authors and source.
• Link to the license and identify this package's modifications.
• Use the material only for non-commercial purposes.
• Distribute adaptations under the same license.

The original authors do not endorse this derived package.

Original Source

• Paper: Camouflaged Object Detection, CVPR 2020
• Official code and dataset information: DengPingFan/SINet

Citation

@inproceedings{fan2020camouflaged,
  title     = {Camouflaged Object Detection},
  author    = {Fan, Deng-Ping and Ji, Ge-Peng and Sun, Guolei and Cheng, Ming-Ming and Shen, Jianbing and Shao, Ling},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2020}
}

Repacked at UTC: 2026-07-07T05:03:34.076047+00:00