Skip to content

Latest commit

 

History

History
47 lines (37 loc) · 2.6 KB

DATA.md

File metadata and controls

47 lines (37 loc) · 2.6 KB
Raw

Dataset preparation

In order to reproduce the results in the paper smoothly, the dataset needs to be ingested correctly. Here we identify two modality of datasets

Video dataset

We use class GluonCVMotionDataset to represent each video-based dataset, for example MOT17, TAO-person, AOT, etc.

In order to ingest the original video dataset to GluonCVMotionDataset format, we provide ingestion scripts in data/ingestion folder, please follow the examples to ingest the video datasets.

To ingest your own dataset, organize it in the following structure:

${dataset_root}
| -- raw_data

All artifacts related to the raw dataset are put to the raw_data folder. After ingestion, the dataset structure is expected to be like the following:

${dataset_root}
|-- annotation
  |-- anno.json
  |-- splits.json
|-- cache
|-- raw_data

We also provide the following ingested datasets (anno.json and splits.json files are provided), Please make sure that cfg.DATASETS.ROOT_DIR in the configuration has been pointed to dataset_root.

  • MOT17: MOT17 videos with all 3 set of detections (DPM, FRCNN, SDP) Ingested annotation
  • MOT17_DPM: MOT17 videos with DPM detection Ingested annotation
  • TAO: TAO-person dataset
  • CRP: Caltech Roadside Pedestrains dataset
  • AOT: AOT dataset for airbone object detection and tracking

In order to train with the above ingested datatset, the raw videos need to be downloaded the original data page, and extract them into raw_data folder.

Image dataset

We use class COCO to represent each image-based dataset, for example, COCO17, CrowdHuman. Please follow the example to ingest the image-based datasets.

We provide the following ingested person detection datasets: