CVO Dataset

CVO dataset is a large optical flow dataset, consisting of synthetic video sequences and comprehensive optical flows (including adjacent frame and cross-frame flows).

Dataset Details

• Video Information:
  • Resolution: 512x512
  • FPS: 30
  • Duration: 0.2s
• Number of Training Videos: 11406
• Number of Testing Videos: 536

Data Format

We dump all the data in LMDB format, where we can index keys for correspoinding values. For each data sample, we provide five data groups as follows:

• imgs: List of video frames, [I0, I1, I2, ..., I6].
• imgs_blur: List of blurred video frames, random motion blur is added.
• fflows: List of forward long-range optical flows, [F02, F03, ..., F06].
• bflows: List of backward long-range optical flows, [F20, F30, ...,F60].
• delta_fflows: List of forward local optical flows, [F01, F12, ..., F56].
• delta_bflows: List of backward local optical flows, [F10, F21, ..., F65].

Here shows the structure of the training file:

CVO_train.lmdb
├── __samples__: [0,1,2,...,11405]
├── 00000_imgs: [...] # 7 RGB frames
├── 00000_imgs_blur: [...] # 7 RGB frames (random blurred)
├── 00000_fflows: [...] # 5 optical flows (cross-frame)
├── 00000_bflows: [...] # 5 optical flows (cross-frame)
├── 00000_delta_fflows: [...] # 6 optical flows (adjacent frame)
├── 00000_delta_bflows: [...] # 6 optical flows (adjacent frame)
├── 00001_xxx
...
├── 00002_xxx
...
...
├── 11405_xxx
...
├── __valid_keys__: [imgs, imgs_blur, ..., delta_bflows]
└── __keys__: [...] # All the data keys in this file

The test file CVO_test.lmdb shares similar structure. All the arrays in a list is concatenated along the channel dimension.

Download Links:

Curretly, we only provide Baidu Netdisk links, other platforms will be included later

Split Compressed files:

CVO-test (Extracting Code: 7w27)

CVO-train (Extracting Code: 9s6c)

HuggingFace Site

Uncompressed LMDB files:

CVO-test (Extracting Code: vii9)

CVO-train (Extracting Code: akep)