An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection

22 Apr 2019Youngwan LeeJoong-won HwangSangrok LeeYuseok BaeJongyoul Park

As DenseNet conserves intermediate features with diverse receptive fields by aggregating them with dense connection, it shows good performance on the object detection task. Although feature reuse enables DenseNet to produce strong features with a small number of model parameters and FLOPs, the detector with DenseNet backbone shows rather slow speed and low energy efficiency... (read more)

PDF Abstract

Evaluation results from the paper


Task Dataset Model Metric name Metric value Global rank Compare
Object Detection COCO Mask R-CNN + VoVNet-57(single-scale) Bounding Box AP 42.0 # 20
Real-Time Object Detection COCO RefineDet320 + VoVNet-57 MAP 33.9 # 4
Real-Time Object Detection COCO RefineDet320 + VoVNet-57 FPS 21.2 # 6
Object Detection COCO RefineDet512 + VoVNet-57 (multi-scale) Bounding Box AP 43.6 # 14
Object Detection COCO RefineDet512 + VoVNet-57 (sinlge-scale) Bounding Box AP 39.2 # 29
Object Detection COCO RefineDet320 + VoVNet-57 backbone Bounding Box AP 33.9 # 36
Real-Time Object Detection PASCAL VOC 2007 DSOD300 + VoVNet-27-slim MAP 74.8% # 4
Real-Time Object Detection PASCAL VOC 2007 DSOD300 + VoVNet-27-slim FPS 71 # 1
Object Detection PASCAL VOC 2007 DSOD300 + VoVNet-27-slim MAP 74.8% # 12