Approach

Our solution is a very standard object detection workflow. We experimented with a new SOTA model that was launched whilst the competition was going on.

• Model: We make use of rfdetr, a new SOTA real-time object detection model released by roboflow, that supports ONNX deployment natively.
• Data preparation(data-preparation.ipynb):We split the dataset into train and validation sets using StratifiedGroupKFold with k of 10. We use fold 0 for validation and the other folds for training. We convert our split dataset into COCO format as required by the rfdetr
• Training(rfdetr-training.ipynb): We trained the rfdetr-large variant for 9 epochs using the following configuration: epochs: 9, batch size: 2, grad_accum_steps: 8, lr: 1e-4, resolution: 448
• Inference(rfdetr-inference.ipynb): RfDetr produces 2 checkpoints: regular weights and EMA-based weights. We made use of the EMA-based weights as those have better performance locally and on the leaderboard. We used a low confidence threshold of 0.02
• Ensembling(wbf.ipynb): We use wbf on the output from the inference notebook. We made use of the ensemble-boxes framework with the following parameters:conf_type: 'max, skip_box_thr: 0.0001, iou_thr: 0.6

Everything is explained in more details in the repo below:

Code Repo: https://github.com/DurojaiyeAbisoye/zindi-amini-cocoa-detection-challenge-4th-place-solution

PS: Thanks to others for also posting their solutions we've learned a couple of new approaches.