DAL Shemagh Detection Challenge
Two-Stage Object Detection Pipeline · Team Sekiro
An object detection competition on Kaggle with almost no labels to start from. Team Sekiro's two-stage pipeline scored 0.863 on the private leaderboard, within 0.03 of first place, for 3rd of more than 100 teams.
0.863
Private Leaderboard Score
5 Images
Labeled Positives at the Start
3rd
Of 100+ Teams
Approach
- Propose: YOLO11m detects candidate bounding boxes.
- Filter: a ConvNeXt-Tiny classifier checks each candidate and removes false positives.
- Merge: Weighted Boxes Fusion (WBF) combines overlapping candidates into one box.
Training from five labels
The training data started with only 5 labeled positive images. An iterative pseudo-labeling loop on the unlabeled imagery grew it round by round: confident predictions became new training labels, at no extra annotation cost. HSV photometric augmentation made the models robust to lighting and color changes.
Result
0.863 on the private leaderboard, within 0.03 of 1st place, for 3rd of 100+ teams. The full 3rd-place solution write-up is published on Kaggle.