[Paper Review] DetNet: A Backbone network for Object Detection

Zeming Li, Chao Peng+4 more|arXiv (Cornell University)|Apr 17, 2018

Advanced Neural Network ApplicationsComputer Science4 references244 citations

TL;DR

DetNet introduces a backbone designed specifically for object detection that maintains high spatial resolution in deeper layers with dilated bottleneck blocks, yielding state-of-the-art COCO results with lower FLOPs than comparators.

ABSTRACT

Recent CNN based object detectors, no matter one-stage methods like YOLO, SSD, and RetinaNe or two-stage detectors like Faster R-CNN, R-FCN and FPN are usually trying to directly finetune from ImageNet pre-trained models designed for image classification. There has been little work discussing on the backbone feature extractor specifically designed for the object detection. More importantly, there are several differences between the tasks of image classification and object detection. 1. Recent object detectors like FPN and RetinaNet usually involve extra stages against the task of image classification to handle the objects with various scales. 2. Object detection not only needs to recognize the category of the object instances but also spatially locate the position. Large downsampling factor brings large valid receptive field, which is good for image classification but compromises the object location ability. Due to the gap between the image classification and object detection, we propose DetNet in this paper, which is a novel backbone network specifically designed for object detection. Moreover, DetNet includes the extra stages against traditional backbone network for image classification, while maintains high spatial resolution in deeper layers. Without any bells and whistles, state-of-the-art results have been obtained for both object detection and instance segmentation on the MSCOCO benchmark based on our DetNet~(4.8G FLOPs) backbone. The code will be released for the reproduction.

Motivation & Objective

• Identify limitations of ImageNet-classification backbones for detection tasks.
• Design a backbone that preserves spatial resolution without sacrificing receptive field.
• Demonstrate DetNet’s effectiveness when used with detection heads like FPN for COCO object detection and instance segmentation.

Proposed method

• Introduce extra stages (e.g., P6) into the backbone while fixing downsampling to keep 16x downsampling after stage 4.
• Use dilated bottleneck blocks with 1x1 projection to create new stages without increasing spatial resolution cost.
• Maintain the same stage count as the detector (e.g., FPN) to ensure pretraining on ImageNet remains compatible.
• Benchmark DetNet-59 (based on ResNet-50) within an FPN framework against ResNet backbones on COCO.
• Compare DetNet variants trained from ImageNet pretraining versus training from scratch to isolate backbone effects.

Experimental results

Research questions

• RQ1 Does a detection-optimized backbone that preserves high-resolution deep features improve localization accuracy and small/large object detection on COCO?
• RQ2 Can DetNet-59 outperform ResNet-50/FPN baselines with lower or comparable FLOPs while enabling better detection and instance segmentation results?

Key findings

• DetNet-59 achieves higher mAP and AP metrics than ResNet-50 when paired with FPN (e.g., mAP improves from 37.9 to 40.2 in their ablations).
• DetNet-59 with FPN outperforms ResNet-101-based backbones in COCO detection despite having fewer FLOPs (4.8G vs 7.6G).
• DetNet-59 from scratch (trained without ImageNet weights) still surpasses ResNet-50 from scratch in COCO FPN results (36.3 vs 34.5 mAP).
• DetNet-59 shows notable gains for large objects (APl) and higher AR at IoU=85, indicating improved localization.
• DetNet-59-based Mask R-CNN achieves state-of-the-art instance segmentation results on COCO test-dev compared with several ResNet-101 baselines.