We require the participator to submit the results as a single .zip file. Each .txt file in the .zip file contains the results of the corresponding image. Notably, the results of each image must be stored in the archive's root folder.
The results file for each task should be stored in the SAME format as the provided ground-truth file, i.e., the JSON text-file containing one object instance per line. If there exists no output detection result, please provide an empty file. We suggest the participator reviewing the ground truth format before proceeding. For different tasks, each line in the text-file contains different content. Each text file stores the detection results of the corresponding image, with each line containing an object instance in the image. The format of each line is as follows:
|1||[bbox_left]||The x coordinate of the top-left corner of the predicted bounding box|
|2||[bbox_top]||The y coordinate of the top-left corner of the predicted object bounding box|
|3||[bbox_width]||The width in pixels of the predicted object bounding box|
|4||[bbox_height]||The height in pixels of the predicted object bounding box|
|5||[score]||The score in the DETECTION result file indicates the confidence of the predicted bounding box enclosing an object instance|
|6||[object_category]||The object category indicates the type of annotated object|
|7||[truncation]||The score in the GROUNDTRUTH file indicates the degree of object parts appears outside a frame|
|8||[occlusion]||The score in the GROUNDTRUTH file indicates the fraction of objects being occluded|
For Task 1 (i.e., object detection in images), we mainly focus on human and vehicles in our daily life, and define ten object categories of interest including pedestrian, person, car, van, bus, truck, motor, bicycle, awning-tricycle, and tricycle. Notably, if a human maintains standing pose or walking, we classify it as pedestrian; otherwise, it is classified as a person. Some rarely occurring specific objects (e.g., machineshop truck, forklift truck, and tanker) are ignored in evaluation. Meanwhile, if the truncation ratio of the object is larger than 50%, it is skipped during evaluation. To obtain results on the VisDrone2019 test-challenge set, the participators must generate the results in default format (see here) and uploaded to the evaluation server.
We require each evaluated algorithm to output a list of detected bounding boxes with confidence scores for each test image in the predefined format. Please see the results format for more detail. Similar to the evaluation protocol in MS COCO , we use AP, APIOU=0.50, APIOU=0.75, ARmax=1, ARmax=10, ARmax=100, and ARmax=500 metrics to evaluate the results of detection algorithms. Unless otherwise specified, the AP and AR metrics are averaged over multiple intersection over union (IoU) values. Specifically, we use ten IoU thresholds of [0.50:0.05:0.95]. All metrics are computed allowing for at most 500 top-scoring detections per image (across all categories). These criteria penalize missing detection of objects as well as duplicate detections (two detection results for the same object instance). The AP metric is used as the primary metric for ranking the algorithms. The metrics are described in the following table.
The above metrics are calculated over object categories of interest. For comprehensive evaluation, we will report the performance of each object category. The evaluation code for object detection in images is available on the PANDA github.
|AP||100%||The average precision over all 10 IoU thresholds (i.e., [0.5:0.05:0.95]) of all object categories|
|APIOU=0.50||100%||The average precision over all object categories when the IoU overlap with ground truth is larger than 0.50|
|APIOU=0.75||100%||The average precision over all object categories when the IoU overlap with ground truth is larger than 0.75|
|ARmax=1||100%||The maximum recall given 1 detection per image|
|ARmax=10||100%||The maximum recall given 10 detections per image|
|ARmax=100||100%||The maximum recall given 100 detections per image|
|ARmax=500||100%||The maximum recall given 500 detections per image|
 T. Lin, M. Maire, S. J. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. L. Zitnick, “Microsoft COCO: common objects in context,” in Proceedings of European Conference on Computer Vision, 2014, pp. 740–755