ljl02521 / self-made_ava_dataset_tool

Some scripts on generating homemade AVA format datasets

Installation：

mmcv: pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/{cu_version}/{torch_version}/index.html Please replace {cu_version} and {torch_version} in the url to your desired one.

  For example: pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.9.0/index.html

mmdet: pip install mmdet

Use:

Step 1: python detectron2_outvia3.py path/to/your/faster_rcnn_r50_fpn_2x_coco.py / path/to/your/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth / --input path/to/your/keyframes//.jpg / --gen_via3 --output path/to/your/annotations_proposal --score-thr 0.5 --show

    For example:
    python detectron2_outvia3.py E:\ava\faster_rcnn_r50_fpn_2x_coco.py 
    E:\ava\faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth
    --input E:\ava\org_img\*.jpg
    --gen_via3
    --output E:\ava\annotations_proposal
    --score-thr 0.5
    --show
    
    faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth can be download here:
    https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth

You can get a JSON file containing the coordinate box and score. For example: self-made_ava_dataset_tool/annotations_proposal/org_img_proposal.json

Step2: Make annotation files in AVA dataset V2.1 format. For example: self-made_ava_dataset_tool/ann_csv/ground_truth.csv Via3 can be download here: https://www.robots.ox.ac.uk/~vgg/software/via/. Open the via3/via_image_annotator.html.Then, upload the image and the JSON file generated in step 1. Then You can then adjust candidate boxes and delete unwanted candidates while tagging. Save ground truth json,and run gt_json2csv.py. The aciton kinds you can change in detectron2_outvia3.py.

Step3(optional)：slowfast needs predicted boxes. You can run predicted_json2csv.py to get it. In this step,You just delete the proposals you don't need, and you don't have to resize or reposition them. For example: self-made_ava_dataset_tool/ann_csv/predicted_ann.csv