Datasets:
| # Crop parameters | |
| One square crop box per clip that frames the signer, found by running a YOLO person detector on every frame | |
| (code: [pjm-data, step 05](https://github.com/crocosign/pjm-data/tree/main/dataset-processing/05_crop_params)). The videos are stored uncropped; the crop is | |
| applied when loading data. | |
| - `crop_params.lmdb/`: LMDB with the clip key (the `__key__` of the segment record) as key and a msgpack map as | |
| value: `x_start`, `y_start`, `x_end`, `y_end` and `padding` (`top`, `bottom`, `left`, `right`). It covers all | |
| 39,813 segments, before the dataset filters. | |
| - `failed.csv`: the 4 clips that could not be opened (`video_name`, `exception`). | |
| ```python | |
| import lmdb, msgpack | |
| env = lmdb.open("crop_params.lmdb", readonly=True, lock=False) | |
| with env.begin() as txn: | |
| box = msgpack.unpackb(txn.get(b"001_000_M")) | |
| ``` | |