KlingTeam/Context-as-Memory-Dataset

<div align="center"> <h1>Context as Memory: Scene-Consistent Interactive Long Video Generation with Memory Retrieval</h1> <h1>SIGGRAPH Asia 2025</h1> <p> <a href="https://context-as-memory.github.io/">[Project page]</a> <a href="https://arxiv.org/pdf/2506.03141">[ArXiv]</a> <a href="https://huggingface.co/datasets/KwaiVGI/Context-as-Memory-Dataset">[Dataset]</a> </p> </div> # File Structure To prepare the dataset for use, merge the parts into a single zip file using the following command: ```bash cat Context-as-Memory-Dataset_* > Context-as-Memory-Dataset.zip ``` After extracting `Context-as-Memory-Dataset.zip`, the dataset will be organized as follows: ``` Context-as-Memory-Dataset ├── frames │ ├── AncientTempleEnv_0 │ │ ├── 0000.png │ │ ├── 0001.png │ │ ├── 0002.png │ │ └── ... │ ├── AncientTempleEnv_1 │ │ ├── 0000.png │ │ ├── 0001.png │ │ ├── 0002.png │ │ └── ... │ └── ... │ ├── jsons │ ├── AncientTempleEnv_0.json │ ├── AncientTempleEnv_1.json │ └── ... │ ├── overlap_labels │ ├── AncientTempleEnv_0 │ │ ├── 0.json │ │ ├── 1.json │ │ ├── 2.json │ │ └── ... │ ├── AncientTempleEnv_1 │ │ ├── 0.json │ │ ├── 1.json │ │ ├── 2.json │ │ └── ... │ └── ... │ └── captions.txt ``` # Explanation of Dataset Parts - **`frames/`**: 100 subdirectories, each containing 7,601 video frame images. - **`jsons/`**: 100 JSON files, each storing the camera pose (position + rotation) of every frame in the corresponding long video. - **`overlap_labels/`**: 100 subdirectories, each containing 7,601 JSON files, where each file records the indices of overlapping frames corresponding to that frame. - **`captions.txt`**: Captions annotated for a segment of a long video, from a given starting frame to an ending frame. - We also provide a simple code file, `tools.py`, which can convert (x, y, z, yaw, pitch) into RT, and can also select a specific frame as the reference frame to align the RT of other frames to its coordinate system.