| Submission name |
ctl-maskrcnn
|
| Submission time (UTC) |
Sept. 23, 2025, 1:10 p.m.
|
| User |
anas.gouda
|
| Task |
Model-free 2D detection of unseen objects |
| Dataset |
HOPEv2 |
| Description |
Centroid Triplet Loss identification method from https://arxiv.org/abs/2404.06277 but uses maskrcnn for segmentation instead of sam. Maskrcnn was trained on Nvidia falling things and DoPose.
|
| Evaluation scores |
| AP: | 0.385 |
| AP50: | 0.573 |
| AP75: | 0.434 |
| AP_large: | 0.447 |
| AP_medium: | 0.064 |
| AP_small: | 0.089 |
| AR1: | 0.406 |
| AR10: | 0.440 |
| AR100: | 0.440 |
| AR_large: | 0.513 |
| AR_medium: | 0.063 |
| AR_small: | 0.089 |
| average_time_per_image: | 0.534 |
|