The Multiview Moving Camera Dataset (MMCD)

Dataset Summary

To our knowledge, no available dataset supports multiple video stitching with N > 2 freely moving cameras. Thus, we created MMCD, a new multiview dataset. MMCD contains 53 scenes, each with 3–6 approximately synchronized handheld videos captured on diverse mobile devices under varied environmental and lighting conditions. Camera models, resolutions, and motions vary substantially. Overall, MMCD contains 221 videos and 47,040 frames.

To facilitate quantitative evaluation, we manually annotated keypoints in the overlap regions of the first frames and propagated them through the sequences using CoTracker3. The annotations are used solely for evaluation.

The dataset is provided as a single file, MMCD.zip. Unzipping it gives the following structure:

MMCD/
├── frames/
│   └── <scene>/
│       ├── video0/
│       ├── video1/
│       └── ...
└── kps_annotations/
    └── <scene>/
        ├── kps_<scene>.pt
        └── pred_visibility_<scene>.pt

N - the number of videos in the scene.

Data Fields

In addition to the frames folder, each scene has 2 tensors, one for the keypoint annotations and one for the visibility flags.

• frames/<scene>/video<i>/: the frames of video i of the scene.
• kps_annotations/<scene>/kps_<scene>.pt: a PyTorch tensor of shape [num_frames, N, keypoint_index, 2], with the 2D position of every keypoint in every frame of every video, in pixels of the original frames.
• kps_annotations/<scene>/pred_visibility_<scene>.pt: a PyTorch tensor of shape [num_frames, N, keypoint_index], with the visibility flags predicted by CoTracker3: 1 if the keypoint is visible in that video's frame, and 0 otherwise.

Data Statistics

MMCD has no train/test splits; all of the scenes are used for evaluation. It is divided into 4 categories based on N, the number of cameras:

	N = 3	N = 4	N = 5	N = 6	Total
Scenes	22	10	11	10	53

All of the videos are 30 fps.

N = 3  (22 scenes)

Scene	Length	Resolution (H×W)
scene_1	9 sec	478×848
scene_2	6 sec	478×848
scene_4	7 sec	478×848
scene_6	8 sec	478×848
scene_7	9 sec	478×848
scene_8	10 sec	478×848
scene_9	6 sec	478×848
scene_10	7 sec	478×848
scene_12	8 sec	478×848
scene_13	9 sec	478×848
scene_15	10 sec	478×848
scene_18	7 sec	478×848
scene_19	8 sec	478×848
scene_20	9 sec	478×848
scene_21	10 sec	478×848
scene_29	6 sec	480×840
scene_39	10 sec	480×840
scene_67	10 sec	480×840
scene_69	9 sec	480×840
scene_70	10 sec	480×840
scene_106	5 sec	480×840
scene_110	9 sec	1200×800

N = 4  (10 scenes)

Scene	Length	Resolution (H×W)
scene_25	5 sec	480×840
scene_30	5 sec	480×840
scene_31	6 sec	480×840
scene_35	6 sec	480×840
scene_47	7 sec	480×840
scene_49	5 sec	480×840
scene_54	8 sec	480×840
scene_64	6 sec	480×840
scene_81	8 sec	480×840
scene_92	10 sec	450×478

N = 5  (11 scenes)

Scene	Length	Resolution (H×W)
scene_33	7 sec	480×840
scene_46	8 sec	480×840
scene_51	7 sec	480×840
scene_62	9 sec	480×840
scene_63	7 sec	480×840
scene_65	5 sec	480×840
scene_66	7 sec	480×840
scene_72	8 sec	480×840
scene_76	6 sec	480×840
scene_77	6 sec	480×840
scene_98	6 sec	480×840

N = 6  (10 scenes)

Scene	Length	Resolution (H×W)
scene_27	7 sec	480×840
scene_50	5 sec	480×840
scene_52	6 sec	480×840
scene_53	5 sec	480×840
scene_55	7 sec	480×840
scene_57	6 sec	480×840
scene_58	7 sec	480×840
scene_59	6 sec	480×840
scene_60	5 sec	480×840
scene_101	9 sec	480×840

Personal and Sensitive Information

The dataset does not contain personal or sensitive information.

Citation

@inproceedings{Ifergane:NeurIPS:2026:MViS,
  title     = {Video Stitching from Multiple Moving Cameras},
  author    = {Ifergane, Shira and Amoyal, Roy and Benishay, Shahar and Freifeld, Oren},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}