Reproduce “RetinaMask: Learning to predict masks improves state-of-the-art single-shot detection for free” on OpenResearch

About this paper

Cheng-Yang Fu, Mykhailo Shvets, Alexander C. Berg

RetinaMask, developed by researchers at UNC at Chapel Hill, enhances the RetinaNet single-shot object detector through training-time improvements, including a self-adjusting bounding box regression loss, an improved anchor matching policy, and an instance mask prediction module. This approach enables state-of-the-art detection accuracy, competitive with two-stage detectors, without incurring additional inference cost for detection tasks.

Key insights

• Incorporating an auxiliary task like instance mask prediction during training can significantly boost the primary object detection accuracy of single-shot models without increasing their inference-time computational cost.
• Dynamic adjustment of loss function hyperparameters, such as the Smooth L1 control point, leads to more robust and effective bounding box regression, reducing reliance on manual tuning.
• More inclusive anchor matching policies that ensure all ground truth objects contribute positive examples, even those with minimal overlap, can improve overall detection performance.

Summary by alphaXiv.