facebookresearch / Detectron

FAIR's research platform for object detection research, implementing popular algorithms like Mask R-CNN and RetinaNet.
Apache License 2.0
26.27k stars 5.46k forks source link

evaluate_box_proposals is not used. How to draw Recall-IOU curve for coco and pascal voc? #436

Open arasharchor opened 6 years ago

arasharchor commented 6 years ago

Expected results

I expected to see recall-iou curve for coco and pascal voc. However, it seems "evaluate_box_proposals" function which is for drawing this curve is not used in both dataset evaluation. In RPN_ONLY mode also "generate_rpn_on_range٬ in test_engine.get_eval_functions does not have box_proposal evaluation. In "task_evaluation" function, the task "box_proposal" has been been used.

Actual results

no curve is output. gt_overlaps and num_pos are empty.

Detailed steps to reproduce

I put the following commands in the beginning of task_evaluation.evaluate_all(

    roidb = dataset.get_roidb()
    logger.info("Evaluate Proposals")
    evaluate_box_proposals(dataset, roidb)
    logger.info('Evaluating proposals is done!')

and run

python2 tools/test_net.py --cfg configs/getting_started/tutorial_1gpu_e2e_faster_rcnn_R-101-FPN.yaml TEST.WEIGHTS output/tutorial_1gpu_e2e_faster_rcnn_R-101-FPN-dota-voc/train/voc_2007_train/generalized_rcnn/model_final.pkl NUM_GPUS 1

System information

mamunir commented 6 years ago

@smajida are u able to solve this issue? please share