For Task 2 (i.e., zero-shot object detection in images), we mainly focus detect objects of unavailable categories during training from remote sensing images. For the GZSD setting, test images contain both the seen and unseen categories. To obtain results on the VisDrone2023 test-GZSD, the participators must generate the results in default format (see here) and upload 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. we use mAP and Recall@100 metrics to evaluate the results of detection algorithms. Unless otherwise specified, the mAP and Recall metrics are averaged over 0.5 intersection over union (IoU) values. Specifically, we use Seen mAP, Seen Recall@100, Unseen mAP and Unseen Recall@100 to evaluate the results of detection algorithms. Furthermore, we also utilize Harmonic Mean (HM) to evaluate the performance. The HM metric is used as the primary metrics for ranking the algorithms. The metrics are described in the following table.
| Measure | Perfect | Description |
| Seen mAP | 100% | The average precision over 0.5 threshold of all seen object categories |
| Seen Recall@100 | 100% | The average recall over 0.5 threshold of all seen object categories |
| Unseen mAP | 100% | The average precision over 0.5 threshold of all unseen object categories |
| Unseen Recall@100 | 100% | The average recall over 0.5 threshold of all Unseen object categories |
| HM | 100% | The overall mAP performance of seen and unseen categories |