Checklist ROAI 2026 Selection Camp Qualifier · Task 1
Professor Grammar vs. the Box Avalanche
Filter the raw, low-threshold detections of a frozen Faster R-CNN into clean pedestrian boxes, first with vanilla NMS and then with any better filtering strategy.
The task
A frozen COCO-pretrained Faster R-CNN (ResNet-50-FPN) was run on the PennFudanPed pedestrian dataset with a confidence threshold of 0.05, producing many duplicate and false-positive boxes. Contestants receive these candidate boxes and scores for all images, plus ground-truth boxes for the 60% training split; the 40% test split has no ground truth.
Subtask 1 requires an implementation of vanilla (greedy) non-maximum suppression. Subtask 2 is open-ended: produce the best possible final set of boxes per test image using only the provided inputs. Training object-detection models is prohibited.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- PennFudanPed images (60% training split with ground-truth boxes, 40% test split without), frozen detector predictions (x1, y1, x2, y2, score) for both splits, and a starter kit with loading, IoU and evaluation utilities.
- You submit
- CSV with subtaskID (1 or 2), datapointID (image id) and answer as a list of boxes [[x1,y1,x2,y2], …] or [] if none remain.
- Scoring
- 0.4 · [email protected] + 0.6 · [email protected] with one-to-one matching at IoU ≥ 0.5. Subtask 1 reaches 30 points at about 0.6; subtask 2 gives 30 points at about 0.84 and 70 points at about 0.95.
- Rules
- Training object detectors such as Faster R-CNN is strictly prohibited
- On-site at Genesis College, Bucharest, through the Nitro Contestant Cloud
- Only the packages listed in the round rules may be used
- Only submissions selected by the contestant count towards the final ranking
- Format
- ROAI 2026 qualifier for the selection camp, 8 April 2026 (06:15–11:15 UTC on the judge); single paper for all medallists of both sections.