PaperlayerView exact revision on arXiv

Source-linked AI summary

SqueezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real-Time Object Detection for Autonomous Driving

Bichen Wu, Alvin Wan, Forrest Iandola, Peter H. Jin, Kurt Keutzer

arXiv:1612.01051v4cs.CV

TL;DR

Autonomous-driving object detection must balance accuracy with real-time speed, compact model size, and low energy use. SqueezeDet is a single-stage fully convolutional detector, and the authors report the same accuracy as previous baselines with 30.4x smaller size, 19.7x faster inference, and 35.2x lower energy.

  • Problem

    Autonomous-driving image detectors must balance accuracy with real-time inference, small model size, and low power use for embedded deployment.

    Source: §1 · p. 1

  • Method

    SqueezeDet uses a fully convolutional, single-stage network that jointly proposes regions and classifies them, producing detections in one network forward pass.

    Source: §3.1 · p. 3, §5 · p. 8

  • Results

    30.4x smaller model size, 19.7x faster inference speed, and 35.2x lower energy at the same accuracy as previous baselines.

    Source: §5 · p. 8

  • Takeaways & Limitations

    The authors report that SqueezeDet and SqueezeDet+ meet their design goals of being small, fast, energy efficient, and accurate for autonomous driving.

    Source: §5 · p. 8

Abstract

from arXiv · show

Object detection is a crucial task for autonomous driving. In addition to requiring high accuracy to ensure safety, object detection for autonomous driving also requires real-time inference speed to guarantee prompt vehicle control, as well as small model size and energy efficiency to enable embedded system deployment. In this work, we propose SqueezeDet, a fully convolutional neural network for object detection that aims to simultaneously satisfy all of the above constraints. In our network, we use convolutional layers not only to extract feature maps but also as the output layer to compute bounding boxes and class probabilities. The detection pipeline of our model only contains a single forward pass of a neural network, thus it is extremely fast. Our model is fully-convolutional, which leads to a small model size and better energy efficiency. While achieving the same accuracy as previous baselines, our model is 30.4x smaller, 19.7x faster, and consumes 35.2x lower energy. The code is open-sourced at \url{https://github.com/BichenWuUCB/squeezeDet}.

1. Introduction

Autonomous-driving object detection must be accurate and real-time while fitting embedded systems’ limits on model size and energy. SqueezeDet addresses these requirements with a fully convolutional detection network.

Introduction

  • Autonomous-driving image detectors must provide accuracy, real-time speed, small model size, and energy efficiency to support vehicle control and embedded deployment.The introduction notes that embedded automotive processors have far smaller power budgets than desktop or rack systems.Source: §1 · p. 1
  • SqueezeDet uses stacked convolutions to extract image features, then ConvDet predicts bounding boxes and categories before filtering produces final detections.Its backbone is SqueezeNet, and the complete model remains smaller than 8MB.Source: §1 · pp. 1–2
  • 57.2 FPS at 1242x375 resolution; energy use is 1.4J per image on a TITAN X GPU, about 84X less than Faster R-CNN.Source: §1 · p. 2
  • A trained SqueezeDet model achieved the best average precision across all three cyclist-detection difficulty levels in the KITTI challenge.Source: §1 · p. 2

2. Related Work

Related work contrasts proposal-based detection’s computational cost with single-stage and fully convolutional approaches, alongside research on smaller CNNs and grid-valued outputs.

2.1. CNNs for object detection

  • R-CNN brought substantial accuracy gains over HOG+SVM and DPM leaders, but per-proposal inference ran at ≤1 fps and ≥200 J/frame; later amortization strategies culminated in Faster R-CNN [22].Source: §2.1 · p. 2
  • R-FCN is fully convolutional, achieves accuracy competitive with R-CNN, and amortizes more computation across region proposals.Source: §2.1 · p. 2
  • KITTI’s top-ranked published methods mostly used Faster R-CNN, but reported no real-time inference; YOLO integrated proposal and classification in one fast stage and achieved real-time speed.Source: §2.1 · p. 2

2.2. Small CNN models

  • SqueezeNet uses 4.8MB of parameters (50x smaller than AlexNet) while matching or exceeding AlexNet-level ImageNet accuracy; GoogLeNetv1 uses 53MB and matches VGG-19-level accuracy.Source: §2.2 · p. 2

2.3. Fully convolutional networks

  • FCNs output a grid rather than a vector from the final parameterized layer, supporting pixelwise segmentation and classification via convolutional grids followed by average pooling.Source: §2.3 · pp. 2–3

3. Method Description

SqueezeDet combines a single-stage convolutional detection pipeline with an anchor-based ConvDet layer and end-to-end training, using SqueezeNet variants to target model size and energy efficiency.

3.1. Detection Pipeline

  • The network extracts a low-resolution, high-dimensional feature map, then predicts boxes and class probabilities at grid centers in one pass before top-confidence selection and NMS.Each box has a confidence score reflecting object likelihood and predicted overlap, plus conditional class probabilities.Source: §3.1 · p. 3, Figure 1 · p. 3

3.2. ConvDet

  • With matched feature-map, grid, class, and anchor sizes, 3×3 ConvDet uses ≈0.46×10^6 parameters versus ≈212×10^6 for YOLO’s FcDet, 460× fewer.ConvDet slides across the feature map to predict box coordinates, confidence, and class probabilities for each anchor, unlike Faster R-CNN’s RPN, which leaves classification to later fully connected layers.Source: §3.2 · pp. 3–4

3.3. Training protocol

  • SqueezeDet trains end-to-end with a multi-task loss, assigning each ground-truth box to its highest-IoU anchor so only responsible anchors contribute localization loss.Non-responsible anchors are also penalized for confidence predictions.Source: §3.3 · pp. 4–5

3.4. Neural Network Design

  • SqueezeNet is chosen as the backbone for model size and energy efficiency, using Fire modules with a 1×1 squeeze layer and parallel 1×1 and 3×3 expand layers.The authors use ImageNet-pretrained SqueezeNet v1.1 for SqueezeDet and a 0.75 squeeze-ratio variant for SqueezeDet+, then add two randomly initialized Fire modules before ConvDet.Source: §3.4 · pp. 5–6

4. Experiments

Experiments on KITTI evaluate SqueezeDet’s detection accuracy, recall, speed, and model size against baselines, then explore accuracy–cost tradeoffs across design variations.

4.1. KITTI object detection

  • The evaluation used a half-and-half split of KITTI’s 7,381 training images, with validation-set AP for car, cyclist, and pedestrian detection.Inputs were scaled to 1242x375 by default; models used nine anchors per grid, and inference retained the top 64 detections before NMS.Source: §4.1 · p. 6
  • SqueezeDet+ achieved the highest mAP across classes and difficulty levels and car accuracy on par with Faster-RCNN + VGG16.Both VGG16- and ResNet50-backed ConvDet variations also achieved competitive AP.Source: §4.1 · p. 6, Table 2 · p. 7, Table 3 · p. 7
  • 19x faster: SqueezeDet+ matched Faster-RCNN+VGG16 accuracy at nearly 19 times its inference speed; SqueezeDet reached 57.2 frames per second.The speed benchmark used a TITAN X GPU with batch size 1; the paper describes this as almost twice the 30 FPS real-time standard.Source: §4.1 · p. 7, Table 2 · p. 7
  • 61X smaller: SqueezeDet was 61X smaller than Faster R-CNN + VGG16 and 30X smaller than Faster R-CNN + AlexNet.Replacing VGG16’s fully connected and RPN layers with ConvDet yielded a 57.4MB model; SqueezeDet was also 95X smaller than YOLO.Source: §4.1 · p. 7, Table 2 · p. 7
  • Over 80%: keeping the top 64 predictions gave SqueezeDet and SqueezeDet+ overall recall above 80%; using all boxes raised it to 91% and 92%, respectively.With 1.5X image resolution, SqueezeDet’s maximum recall using all boxes rose to 95%.Source: §4.1 · p. 7, Figure 5 · p. 8

4.2. Design space exploration

  • Scaling up the input image decreased mAP while increasing FLOPs and reducing speed; image scaling also changed the recall ceiling.The 1.5X-resolution model’s maximum recall was 95%, while the 0.75X-resolution model’s was 0.90.Source: §4.2 · p. 8, Figure 5 · p. 8
  • Increasing the number of anchors only modestly increased ConvDet models’ size, FLOPs, and memory footprint.The design-space experiments compared accuracy, FLOPs, inference speed, model size, and activation memory footprint.Source: Table 4 · p. 7, §4.2 · p. 8
  • A more powerful backbone significantly improved accuracy but substantially increased FLOPs, model size, and memory footprint.The exploration compared SqueezeDet, its modifications, and SqueezeDet+.Source: §4.2 · pp. 7–8, Table 4 · p. 7

5. Conclusion

SqueezeDet is a fully convolutional network for real-time object detection that integrates region proposal and classification into ConvDet. Compared with previous baselines, it achieves the same accuracy with a smaller model, faster inference, and lower energy use.

5. Conclusion

  • At the same accuracy as previous baselines, SqueezeDet is 30.4x smaller, 19.7x faster, and consumes 35.2x lower energy.Source: §5 · p. 8
  • SqueezeDet integrates region proposal and classification into ConvDet, which is orders of magnitude smaller than its fully connected counterpart.Source: §5 · p. 8
  • SqueezeDet and SqueezeDet+ are designed to be small, fast, energy efficient, and accurate for autonomous driving.Source: §5 · p. 8

1. Low Power Neural Net Design Guideline

Because off-chip DRAM accesses consume far more energy than SRAM accesses or arithmetic, the guideline focuses on reducing memory accesses by shrinking neural-network memory footprints. With custom hardware and effective scheduling, parameters and activations can remain on-chip, avoiding DRAM accesses.

1. Low Power Neural Net Design Guideline

  • 100x more energy is consumed by a DRAM access than by an SRAM access or floating-point operation, motivating the focus on reducing memory accesses.The paper identifies memory-access reduction as its main focus.Source: §1 · p. 10
  • Reducing model parameters and intermediate activations shrinks the computation’s memory footprint; on CPUs, programmers generally encourage cache locality, while GPU programming can involve manual SRAM management.When required data exceeds on-chip SRAM capacity, processors must access off-chip DRAM.Source: §1 · p. 10
  • With fewer parameters and perfect scheduling on custom hardware, all model parameters and activations of any two consecutive layers can fit in SRAM, eliminating off-chip DRAM accesses and saving energy.Custom hardware permits more granular memory scheduling.Source: §1 · p. 10

2. Memory Footprint

SqueezeDet is designed to fit its parameters and intermediate activations in on-chip SRAM, while its activation memory footprint and FLOPs are lower than those of other models. Fire modules and pooling control activation size, and the final detection output is negligible.

2. Memory Footprint

  • At 7.9MB without compression, SqueezeDet’s parameters may fit in on-chip SRAM for reuse during evaluation.The largest intermediate activation is conv1’s 28.3MB output; fusing convolution with max pooling could down-sample it immediately and reduce the stored activation to about 7MB.Source: §2 · p. 10
  • Fire modules reduce activation size by alternating squeeze layers that compress channels with parallel expand layers that retrieve compressed channel information.Later fire modules have more channels, while max pooling reduces spatial resolution to control activation size; the final ConvDet output encoding thousands of proposals is negligible.Source: §2 · p. 10
  • SqueezeDet has a much lower activation memory footprint and performs fewer FLOPs than other models, leading to better energy efficiency.Table 2 compares SqueezeDet and other models in energy efficiency and related aspects.Source: §2 · p. 10, Table 2 · p. 11

3. Experiments

Experiments measure energy use during object detection and report that SqueezeDet and its variations are energy-efficient compared with prior neural-network detectors.

3. Experiments

  • Energy was measured on 3,741 KITTI images at 1242x375 resolution and batch size 1, sampling GPU power every 0.1 seconds and averaging the middle third of periods at or above 20W.Energy per image is computed from average power and inference speed.Source: §3 · p. 11
  • SqueezeDet consumes 1.4J per image, 84× less than Faster R-CNN + VGG16, while SqueezeDet+ uses 4J per image and is >10X more efficient than Faster R-CNN based methods.The paper reports SqueezeDet+ achieves much better accuracy; overall, SqueezeDet and its variations are described as very energy-efficient compared with previous neural-network-based detectors.Source: §3 · pp. 11–12
  • At 0.75× image resolution, mAP drops by 3 percentage points, while inference speed is 1.6× faster and energy consumption is less than 1J per image.Source: §3 · p. 11
  • VGG16 and ResNet50 convolutional layers combined with ConvDet achieve better energy efficiency than Faster R-CNN based models.Source: §3 · p. 11
  • YOLO’s energy use is 7.3J per frame by direct measurement or 4.2J per frame using the 45FPS reported in, comparable with SqueezeDet+ despite SqueezeDet+ receiving 2X more pixels.Source: §3 · pp. 11–12

Related papers

Loading 1612.01051v4…