Tinier-YOLO: A Real-Time Object Detection Method for Constrained Environments

Deep neural networks (DNNs) have shown prominent performance in the field of object detection. However, DNNs usually run on powerful devices with high computational ability and sufficient memory, which have greatly limited their deployment for constrained environments such as embedded devices. YOLO is one of the state-of-the-art DNN-based object detection approaches with good performance both on speed and accuracy and Tiny-YOLO-V3 is its latest variant with a small model that can run on embedded devices. In this paper, Tinier-YOLO, which is originated from Tiny-YOLO-V3, is proposed to further shrink the model size while achieving improved detection accuracy and real-time performance. In Tinier-YOLO, the fire module in SqueezeNet is appointed by investigating the number of fire modules as well as their positions in the model in order to reduce the number of model parameters and then reduce the model size. For further improving the proposed Tinier-YOLO in terms of detection accuracy and real-time performance, the connectivity style between fire modules in Tinier-YOLO differs from SqueezeNet in that dense connection is introduced and fine designed to strengthen the feature propagation and ensure the maximum information flow in the network. The object detection performance is enhanced in Tinier-YOLO by using the passthrough layer that merges feature maps from the front layers to get fine-grained features, which can counter the negative effect of reducing the model size. The resulting Tinier-YOLO yields a model size of 8.9MB (almost $4\times $ smaller than Tiny-YOLO-V3) while achieving 25 FPS real-time performance on Jetson TX1 and an mAP of 65.7% on PASCAL VOC and 34.0% on COCO. Tinier-YOLO alse posses comparable results in mAP and faster runtime speed with smaller model size and BFLOP/s value compared with other lightweight models like SqueezeNet SSD and MobileNet SSD.