Dynamic loss for one-stage object detectors in computer vision

A dynamic loss (DL) is proposed for the one-stage object detectors in computer vision, as an improved version of the focal loss in the literature. The proposed loss features a second-order item which can efficiently scale the conventional loss during training. A gradient update approach is then presented to employ the DL in the mainstream one-stage YOLO-V2 object detector. Experimental results shows that for the PASCAL VOC 2010 dataset, the mean average precision of the YOLO-v2 detector with the proposed DL is 88.51%, which is 2.6, 1.17 and 0.29% higher than that using the conventional mean square error loss, cross entropy loss and naive focal loss. Compared to the focal loss, the DL increases the training convergence speed of the YOLO-v2 detector by two times.