[Paper Review] Tiny-DSOD: Lightweight Object Detection for Resource-Restricted Usages

Yuxi Li, Jiuwei Li+2 more|arXiv (Cornell University)|Jul 29, 2018

Advanced Neural Network ApplicationsComputer Science27 references80 citations

TL;DR

Tiny-DSOD introduces depthwise dense blocks and a depthwise FPN to achieve high mAP with very small parameter size and FLOPs, outperforming other ultra-efficient detectors on VOC, KITTI, and COCO.

ABSTRACT

Object detection has made great progress in the past few years along with the development of deep learning. However, most current object detection methods are resource hungry, which hinders their wide deployment to many resource restricted usages such as usages on always-on devices, battery-powered low-end devices, etc. This paper considers the resource and accuracy trade-off for resource-restricted usages during designing the whole object detection framework. Based on the deeply supervised object detection (DSOD) framework, we propose Tiny-DSOD dedicating to resource-restricted usages. Tiny-DSOD introduces two innovative and ultra-efficient architecture blocks: depthwise dense block (DDB) based backbone and depthwise feature-pyramid-network (D-FPN) based front-end. We conduct extensive experiments on three famous benchmarks (PASCAL VOC 2007, KITTI, and COCO), and compare Tiny-DSOD to the state-of-the-art ultra-efficient object detection solutions such as Tiny-YOLO, MobileNet-SSD (v1 & v2), SqueezeDet, Pelee, etc. Results show that Tiny-DSOD outperforms these solutions in all the three metrics (parameter-size, FLOPs, accuracy) in each comparison. For instance, Tiny-DSOD achieves 72.1% mAP with only 0.95M parameters and 1.06B FLOPs, which is by far the state-of-the-arts result with such a low resource requirement.

Motivation & Objective

• Design a lightweight object detector optimized for resource-restricted usages (low memory and compute).
• Maintain or improve detection accuracy while minimizing parameters and FLOPs.
• Leverage compact architectural blocks to enable training from scratch and efficient inference.
• Investigate front-end semantic fusion via a lightweight feature pyramid.
• Benchmark against state-of-the-art ultra-efficient detectors across multiple datasets.

Proposed method

• Propose depthwise dense blocks (DDB) to replace dense blocks in DenseNet; two variants DDB-a and DDB-b; choose DDB-b for backbone due to better efficiency and accuracy.
• Introduce depthwise FPN (D-FPN) front-end to fuse semantic information across scales using depthwise convolutions and bilinear interpolation for the upsampling path.
• Build Tiny-DSOD on the SSD framework with DSOD-inspired deep supervision to enable training from scratch.
• Train from scratch on VOC07+12 and evaluate on VOC2007, KITTI, and COCO with data augmentation and online hard example mining.
• Compare against Tiny-YOLO, SqueezeDet, MobileNet-SSD, Pelee, and DSOD baselines in terms of mAP, parameters, and FLOPs.

Experimental results

Research questions

• RQ1 Can a depthwise dense block based backbone combined with a depthwise FPN front-end achieve a favorable accuracy vs. resource trade-off for object detection?
• RQ2 What is the impact of DDB-b versus DDB-a on accuracy and compute under similar parameter budgets?
• RQ3 Does the lightweight front-end (D-FPN) provide a meaningful accuracy boost with minimal FLOPs/params increase across benchmarks?
• RQ4 How does Tiny-DSOD perform against other ultra-efficient detectors on VOC07, KITTI, and COCO in terms mAP, parameter size, and FLOPs?

Key findings

• Tiny-DSOD achieves 72.1% mAP on VOC2007 with 0.95M parameters and 1.06B FLOPs.
• DDB-b backbone outperforms DDB-a at similar resource budgets (e.g., 0.90M params: DDB-b achieves 70.2% mAP vs 63.1% with DDB-a).
• D-FPN adds about 0.03B FLOPs and 0.05M parameters but yields ~1.9% mAP improvement on VOC2007.
• Tiny-DSOD runs at 105 FPS (9.5 ms) on 300x300 input on TitanX, surpassing several lightweight detectors in speed while using far fewer parameters.
• On KITTI, Tiny-DSOD achieves 77.0% mAP with 0.85M params and 4.1B FLOPs, outperforming several compact detectors.
• On COCO test-dev, Tiny-DSOD achieves 23.2 AP@IoU 0.5:0.95 with 1.12B FLOPs and 1.15M parameters, outperforming MobileNet-SSD and PeleeNet among lightweight models and even approaching some full-detectors.