CorridorKeyDataset

This dataset contains 1,079 synthetic green-screen clips, 52,480 frames in total. Each clip includes Alpha, BG, FG, and Input data in .exr format, a preview.jpg, and an entry in manifest.json.

This dataset is currently under a limited release. It is for non-commercial uses only, intended for CorridorKey-related projects and experiments. Do not re-upload or re-distribute the dataset.

Please request access via the Hugging Face UI, then join us on the Corridor Creates Discord and let me (@Neex) know that you would like access.

Structure

manifest.json
Clip_0000/
  preview.jpg
  Alpha/
    Frame_0000.exr
    ...
  BG/
    Frame_0000.exr
    ...
  FG/
    Frame_0000.exr
    ...
  Input/
    Frame_0000.exr
    ...
Clip_0001/
...
Clip_1078/

Frame format

• 2048x2048 pixels, float32, scene-linear. Stored as RGBA with the A channel always 1.
• Alpha/: the ground-truth matte. RGB channels are all the same so you can just read one. Values are in [0, 1] with small overshoot from filtering.
• BG/: the green-screen plate with no subject.
• FG/: the subject only, premultiplied by alpha.
• Input/: FG composited over BG.

Manifest format

One entry per clip:

{
  "name": "Clip_0506",
  "author": "Lucas",
  "frames": 49,
  "md5": "c0280c204c611d1bc7a73faac1a3b7e1",
  "description": "bearded man in jacket jumping and landing",
  "source": "Lucas (27)"
}

• author: who made the clip. Empty where unknown.

• frames: number of frames in the clip.

• md5: MD5 hash of the raw bytes of every EXR file in the clip, concatenated in the order Alpha, BG, FG, Input, each folder in frame order. preview.jpg is not part of the hash. Use it to verify a clip after download and to deduplicate identical clips. In Python:

  import hashlib
  from pathlib import Path
  
  def clip_md5(clip_dir):
      h = hashlib.md5()
      for sub in ("Alpha", "BG", "FG", "Input"):
          for f in sorted(Path(clip_dir, sub).glob("Frame_*.exr")):
              h.update(f.read_bytes())
      return h.hexdigest()
  

• description: a short AI-generated description of the shot. Not to be used as captioning data! I added it to make it easier for humans to find specific clips.

• source: the folder name this clip had before the dataset was reorganized, for reference.

Reading EXR

OpenCV reads these files once OpenEXR support is enabled:

import os
os.environ["OPENCV_IO_ENABLE_OPENEXR"] = "1"
import cv2
alpha = cv2.imread("Clip_0000/Alpha/Frame_0000.exr", cv2.IMREAD_UNCHANGED)[..., 0]