TASLE

# TASLE [](https://debby-0527.github.io/TASLE/) [](https://github.com/67Pluuus/MSLoc) TASLE is a dataset for detecting and temporally localizing AIGC-manipulated segments in long videos. It contains real and manipulated videos, temporal boundaries of manipulated segments, and related semantic annotations. The dataset supports video authenticity detection, manipulated-segment localization, and temporal boundary analysis. ## Dataset Splits | Annotation file | Samples | Description | | --- | ---: | --- | | `train.json` | 11,179 | Default training split: 5,506 manipulated and 5,673 real videos | | `test.json` | 1,293 | Default test split: 631 manipulated and 662 real videos | | `test_add.json` | 1,442 | Optional OOD-extended test split; not used by the default evaluation | | `test_segment_classification.json` | 1,000 | Optional annotations for segment-level classification evaluation | | `train_trace_full_video.json` | 11,179 | Optional full-video temporal localization annotations | The default training and evaluation workflow uses `train.json` and `test.json`. ## Directory Structure Download the repository contents into a directory named `MSLoc_assets` so that the released code can resolve the default paths directly: ```text MSLoc_assets/ ├── data/Tasle-CoT-10K/ │ ├── annos/ # JSON annotations │ └── videos/ # Video files ├── DeMamba/ # Stage-1 pretrained weights and released results └── Trace/ # Stage-2 pretrained weights and released assets ``` ## Annotation Format The main annotation fields are shown below: ```json { "source": "video source", "video_path": "relative/path/to/video.mp4", "duration": 10.0, "type": "real or fake", "annotations": [ { "segment": [1.2, 6.8], "label": "object category", "model": "generation model" } ] } ``` `segment` is expressed in seconds and specifies the start and end times of a manipulated segment. For real videos, `annotations` is usually an empty list. ## Download ```bash pip install modelscope modelscope download --dataset L67plus/TASLE --local_dir ./MSLoc_assets ``` The dataset is large, so please ensure that sufficient disk space is available before downloading it. ## Project Resources - [Project page](https://debby-0527.github.io/TASLE/) - [MSLoc code](https://github.com/67Pluuus/MSLoc) Clone the code into a directory named `MSLoc` and place it beside the downloaded `MSLoc_assets` directory: ```text workspace/ ├── MSLoc/ └── MSLoc_assets/ ``` The directory names are intentional: `MSLoc` contains the source code, while `MSLoc_assets` contains the dataset and released checkpoints. Please refer to the README in the MSLoc repository for environment setup, inference instructions, and the expected file paths. ## License This repository is released under the Apache License 2.0. Third-party source videos remain subject to the licenses and terms of use of their respective original datasets.