CVC-ClinicDB (file mirror)

612 frames extracted from 29 colonoscopy sequences, each with a pixel-level polyp mask (Bernal et al., 2015). Native resolution 384x288.

This is a plain file mirror, not a datasets-format repo: images and masks are stored as files so any path-based dataloader can consume them directly after snapshot_download. The dataset viewer is disabled for that reason.

Layout

CVC-ClinicDB/
  Original/*.png          # 612 RGB frames, 384x288
  Ground Truth/*.png      # 612 binary masks, same stem as their image
class_dict.csv            # background=(0,0,0), polyp=(255,255,255)
metadata.csv              # frame_id, sequence_id, image_path, mask_path

metadata.csv carries the sequence_id column: the 612 frames come from only 29 sequences, so a frame-level random split leaks near-duplicate frames across train and validation. Group by sequence_id for an honest split. Its path columns were rewritten to match this repo's layout.

Usage

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="berkaytrhn/cvc-clinicdb",
    repo_type="dataset",
    local_dir="data/cvc-clinicdb",
    allow_patterns=["CVC-ClinicDB/**", "*.csv"],   # skip the TIF tree
)
# -> data/cvc-clinicdb/CVC-ClinicDB/Original, .../Ground Truth

Source and citation

The official distribution is a .rar behind a JavaScript-rendered Drive link on https://polyp.grand-challenge.org/CVCClinicDB/ ; the PNG rendering here follows the widely used Kaggle mirror balraj98/cvcclinicdb. Released for research use. Cite the original paper, not this mirror:

@article{bernal2015wm,
  title     = {WM-DOVA maps for accurate polyp highlighting in colonoscopy:
               Validation vs. saliency maps from physicians},
  author    = {Bernal, Jorge and Sanchez, F. Javier and
               Fernandez-Esparrach, Gloria and Gil, Debora and
               Rodriguez, Cristina and Vilarino, Fernando},
  journal   = {Computerized Medical Imaging and Graphics},
  volume    = {43},
  pages     = {99--111},
  year      = {2015},
  publisher = {Elsevier}
}