السلام عليكم رمضان مبارك على الجميع
\nFirst, thank you to the organizers for hosting this challenge! My approach relied on treating the problem as a YOLO object detection task from two different angles, heavily optimizing for a channel-swapping quirk in the test data, and ensembling the results.
\nHere is a breakdown of the methodology that achieved 0.857 Public LB / 0.839 Private LB.
\nInstead of relying on a single model architecture, I split the problem into two parallel tracks:
\nface and shemagh.face, shemagh_correct, and shemagh_wrong to provide a stronger signal for the right_place classification.1. Data Preparation & Offline Augmentation
\nshemagh was the minority class, I generated synthetic training samples using offline augmentations (channel shuffle, color jitter, Gaussian blur, rotation, CLAHE) to balance the dataset before training.2. Model Training\nBoth tracks were trained independently on 768×768 images but with different hyperparameters optimized for their specific tasks:
\n3. Test-Time Channel Optimization (The \"Channel Map\")\nI noticed that roughly 50.7% of the test images had swapped BGR channels. To fix this dynamically during inference:
\n4. Inference & TTA\nUsing the corrected channel map, I ran Test Time Augmentation (TTA) across 3 scales (768, 832, 896) with horizontal flipping. This resulted in 6 predictions per image, which were merged using class-aware NMS.
\nright_place = 1 if any shemagh bounding box overlapped with a face bounding box (IoU threshold tuned on the validation set).5. Ensembling with Weighted Boxes Fusion (WBF)
\nI used a two-level ensembling strategy to maximize the strengths of both tracks:
\nright_place flag, I used a union approach: if either track predicted a 1, the final prediction was 1.السلام عليكم رمضان مبارك على الجميع
\nFirst, thank you to the organizers for hosting this challenge! My approach relied on treating the problem as a YOLO object detection task from two different angles, heavily optimizing for a channel-swapping quirk in the test data, and ensembling the results.
\nHere is a breakdown of the methodology that achieved 0.857 Public LB / 0.839 Private LB.
\nInstead of relying on a single model architecture, I split the problem into two parallel tracks:
\nface and shemagh.face, shemagh_correct, and shemagh_wrong to provide a stronger signal for the right_place classification.1. Data Preparation & Offline Augmentation
\nshemagh was the minority class, I generated synthetic training samples using offline augmentations (channel shuffle, color jitter, Gaussian blur, rotation, CLAHE) to balance the dataset before training.2. Model Training\nBoth tracks were trained independently on 768×768 images but with different hyperparameters optimized for their specific tasks:
\n3. Test-Time Channel Optimization (The \"Channel Map\")\nI noticed that roughly 50.7% of the test images had swapped BGR channels. To fix this dynamically during inference:
\n4. Inference & TTA\nUsing the corrected channel map, I ran Test Time Augmentation (TTA) across 3 scales (768, 832, 896) with horizontal flipping. This resulted in 6 predictions per image, which were merged using class-aware NMS.
\nright_place = 1 if any shemagh bounding box overlapped with a face bounding box (IoU threshold tuned on the validation set).5. Ensembling with Weighted Boxes Fusion (WBF)
\nI used a two-level ensembling strategy to maximize the strengths of both tracks:
\nright_place flag, I used a union approach: if either track predicted a 1, the final prediction was 1.