PaperlayerView exact revision on arXiv

Source-linked AI summary

Tiny-DSOD: Lightweight Object Detection for Resource-Restricted Usages

Yuxi Li, Jiuwei Li, Weiyao Lin, Jianguo Li

arXiv:1807.11013v1cs.CV

TL;DR

Many object detectors require substantial computation and model memory, limiting their suitability for low-power and low-end uses such as always-on or battery-powered devices. Tiny-DSOD combines a depthwise-dense-block backbone with a depthwise feature-pyramid front-end in a detector designed for resource-restricted use. Across VOC 2007, KITTI, and COCO, Tiny-DSOD outperforms compared lightweight detectors on accuracy, parameter-size, and FLOPs; on VOC 2007 it achieves 72.1% mAP with 0.95M parameters.

  • Problem

    Many object detectors require substantial computation and model memory, limiting their suitability for low-power and low-end uses such as always-on or battery-powered devices.

    Source: §1 · pp. 1–2

  • Method

    Tiny-DSOD combines a depthwise-dense-block backbone with a depthwise feature-pyramid front-end in a detector designed for resource-restricted use.

    Source: §1 · pp. 2–3

  • Results

    Across VOC 2007, KITTI, and COCO, Tiny-DSOD outperforms compared lightweight detectors on accuracy, parameter-size, and FLOPs; on VOC 2007 it achieves 72.1% mAP with 0.95M parameters.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    The results show a resource–accuracy trade-off for Tiny-DSOD across the three evaluated detection benchmarks.

    Source: §5 · p. 10

Abstract

from arXiv · show

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.

1 Introduction

Object detectors can achieve high accuracy but often require substantial computation and memory, limiting their fit for low-power and low-end uses. Tiny-DSOD combines a depthwise-dense backbone and a depthwise feature-pyramid front-end, and outperforms lightweight detectors across three benchmarks and three metrics.

1 Introduction

  • High computation and large model size constrain low-power detection, since computation can raise power budgets and persistent memory access is costly on low-end devices.The paper identifies always-on and battery-powered low-end devices as examples of the affected settings.Source: §1 · pp. 1–2
  • Existing compact detectors reduce resource demands but retain an accuracy gap; MobileNet-SSD trails SSD by 9.2 percentage points on VOC 2007.The paper frames improving the joint trade-off between FLOPs, memory, and accuracy as an open design opportunity.Source: §1 · p. 2
  • Tiny-DSOD combines depthwise dense blocks with a depthwise FPN to target a better resource–accuracy balance.The DDB backbone retains implicit deep supervision, while the D-FPN fuses semantic information across neighboring scales.Source: §1 · pp. 2–3
  • Tiny-DSOD outperforms state-of-the-art ultra-efficient detectors on accuracy, parameter size, and FLOPs across VOC 2007, KITTI, and COCO; on VOC 2007 it records 72.1% mAP, 0.95M parameters, and 1.06B FLOPs.Source: §1 · pp. 2–3

2 Related Works

Object-detection research includes two-stage and single-stage frameworks, with efficiency efforts targeting models suitable for resource-restricted uses. The paper notes that compact detectors still lag full-sized counterparts in accuracy, motivating a better resource–accuracy trade-off.

State-of-the-arts Object Detection Networks

Lightweight Object Detection Networks

  • Efficient detection work uses smaller backbones and depthwise separable convolution, with MobileNet-SSD reporting 68.0% mAP on VOC 2007 at 5.5M parameters and 1.14B FLOPs.SqueezeNet-based SSD is another compact approach; Pelee uses two-way dense connectivity to reduce computation while retaining accuracy for mobile applications.Source: State-of-the-arts Object Detection… · pp. 3–4
  • On VOC 2007, Tiny YOLO reaches 57.1% mAP and MobileNet-SSD 68.0%, while YOLOv2 and full SSD reach 78.6% and 77.2%, respectively.The paper uses these gaps to motivate a better trade-off between resource use and accuracy.Source: State-of-the-arts Object Detection… · p. 4

3 Method

Tiny-DSOD builds on SSD and DSOD with a DenseNet-like backbone and a lightweight feature-pyramid front-end. Its DDB and D-FPN designs aim to reduce computation while retaining dense connections and transferring semantic information across feature scales.

3 Method

  • Tiny-DSOD is based on SSD and DSOD and separates its detector into backbone and front-end parts.Source: §3 · p. 4

3.1 Depthwise Dense Blocks Based Backbone

  • The backbone uses a DenseNet-like structure with depthwise separable convolution in its dense blocks, forming the depthwise dense block (DDB).The design is inspired by DSOD’s training-from-scratch approach for relatively smaller training sets.Source: §3.1 · p. 4
  • DDB-a expands input channels by w, applies depthwise convolution, projects to growth rate g, and concatenates input and output features.The figure distinguishes DDB-a, parameterized by g and w, from DDB-b, parameterized by g.Source: §3.1 · p. 4, Figure 1 · p. 5
  • DDB-a’s stacked complexity grows as O(L^3g^2), requiring small growth rate g and risking reduced discriminative power; adjacent 1×1 projections may also be redundant.Source: §3.1 · p. 4
  • DDB-b has stacked complexity O(L^2g^2), below DDB-a’s O(L^3g^2), and is selected as the backbone unit after reported accuracy and efficiency advantages under similar resource constraints.DDB-b compresses input channels to growth rate g, applies depthwise convolution, and concatenates its output directly to the input without an extra 1×1 projection.Source: §3.1 · pp. 4–5
  • The backbone has four DDB stages and transition layers that fuse channel-wise information and compress channels for efficiency.Growth rate is smaller in shallower, larger-spatial-size stages and increases linearly in deeper stages to save computation.Source: §3.1 · p. 5

3.2 Depthwise FPN based Front-end

  • D-FPN redirects information from deeper, smaller feature maps to shallower ones to address their lack of semantic information.Its reverse path uses bilinear upsampling followed by depthwise convolution, and the resulting maps are added to same-sized lower-layer maps.Source: §3.2 · pp. 5–6
  • D-FPN’s lightweight front-end is reported to boost detection accuracy with a slight increase in computation.Source: §3.2 · p. 6

4 Experiments

Experiments evaluate Tiny-DSOD’s design choices and benchmark performance on PASCAL VOC 2007, KITTI, and COCO, alongside implementation and runtime details. The results show resource-efficient detection across these settings, with trade-offs examined in the backbone and front-end.

4.1 Implementation Details

  • Training uses Caffe and SGD from scratch, with most training strategies following DSOD, including data augmentation, L2 normalization, Smooth L1 localization loss, cross-entropy classification loss, and online hard example mining.The models are trained on a server with a PASCAL TitanX GPU.Source: §4.1 · p. 6

Design Space Exploration

  • With parameters fixed at 0.90M, DDB-b achieves 70.2% mAP versus DDB-a’s 63.1% and uses 1.03B versus 1.68B FLOPs; the authors select DDB-b as standard.Increasing backbone growth rate improves accuracy, but higher growth in shallow stages raises computation cost; the selected G/32-48-64-80 setting offers comparable accuracy with least FLOPs under the same model-size constraint.Source: Design Space Exploration · p. 7
  • D-FPN adds 1.9% performance gain with increases of 0.03B FLOPs and 0.05M parameters.The authors describe these resource increases as tolerable relative to the accuracy gain.Source: Design Space Exploration · p. 7
  • At 300×300 input, Tiny-DSOD runs at 105 fps on a TitanX GPU with batch size 8, faster than all compared ultra-efficient detectors except Tiny-YOLO.The paper attributes the slower-than-Tiny-YOLO result partly to Caffe’s less efficient depthwise convolution implementation and the lack of additional optimization.Source: Design Space Exploration · p. 7

4.3 Benchmark Results on PASCAL VOC2007

  • 72.1% mAP on VOC2007, Tiny-DSOD uses 1/6 the parameters and 1/5 the FLOPs of DSOD-smallest, while Faster R-CNN is 1.1% more accurate but over 140× larger.The paper reports that Faster R-CNN also requires 180× more theoretical computing cost and is practically 10× slower in fps.Source: Table 3 · p. 8, §4.3 · p. 8

4.4 Benchmark Results on KITTI

  • 77.0% mAP on KITTI validation, Tiny-DSOD slightly exceeds SqueezeDet’s 76.7% while using over 50% fewer parameters and FLOPs and running at 15 ms per image.Tiny-DSOD also achieves the highest accuracy in the cars category.Source: Table 4 · p. 9, §4.4 · p. 9

4.5 Benchmark Results on COCO

  • 23.2% mAP on COCO test-dev, Tiny-DSOD outperforms MobileNet-SSD v1 and v2, PeleeNet, and YOLOv2; YOLOv2 has 58× larger model size and 15.6× more FLOPs.The reported metric is AP@IOU[0.5 : 0.95].Source: §4.5 · pp. 9–10, Table 5 · p. 10

5 Conclusion

Tiny-DSOD uses depthwise dense blocks and depthwise feature-pyramid networks to balance resource use and accuracy. Across PASCAL VOC 2007, KITTI, and COCO, it outperforms the compared lightweight detectors on accuracy, FLOPs, and parameter size.

5 Conclusion

  • Tiny-DSOD outperforms the compared lightweight detectors on accuracy, FLOPs, and parameter size across PASCAL VOC 2007, KITTI, and COCO.The comparisons include MobileNet-SSD (v1 and v2), SqueezeDet, and Pelee.Source: §5 · p. 10
  • 72.1% mAP on PASCAL VOC 2007 uses 0.95M parameters and 1.14B FLOPs.The conclusion describes this as a state-of-the-art result at such a low resource requirement.Source: §5 · p. 10
  • Depthwise dense blocks (DDB) and depthwise feature-pyramid networks (D-FPN) are the two blocks used to improve the resource–accuracy trade-off.The paper reports verifying the blocks and detectors through extensive ablation studies.Source: §5 · p. 10

Related papers

Loading 1807.11013v1…