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
TL;DR
Autonomous-driving object detectors must combine accuracy with real-time speed, compact size, and low energy use for embedded deployment. SqueezeDet addresses these constraints with a fully convolutional, single-pass detector, achieving comparable accuracy with 30.4x smaller models, 19.7x faster inference, and 35.2x lower energy.
Problem
Autonomous-driving object detection needs high accuracy, real-time inference, compact models, and low power for safe, deployable embedded systems.
Method
SqueezeDet uses a fully convolutional single-pass network that integrates region proposal and classification into convolutional detection layers.
Results
30.4x smaller, 19.7x faster, and 35.2x lower energy than previous baselines while achieving the same accuracy.
Takeaways & Limitations
SqueezeDet provides an accurate, compact, fast, and energy-efficient object detector designed for autonomous-driving deployment.
Takeaways & Limitations
The SqueezeDet+ experiment uses input images containing twice as many pixels as the YOLO inputs, complicating direct comparison.
Abstract
from arXiv · showhide
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
SqueezeDet targets autonomous-driving object detection by jointly addressing accuracy, real-time speed, small model size, and low power for embedded deployment. It uses a fully convolutional, single-pass pipeline with convolutional prediction layers to generate and filter detections.
- Autonomous-driving perception must detect cars, pedestrians, cyclists, road signs, and other objects accurately in real time for safe control decisions.
- Image-based detection is nearly irreplaceable because cameras are cheaper than LIDAR and produce more abundant, easier-to-annotate data.
- Autonomous-driving detectors must combine high recall and precision with real-time inference and deployment on low-power embedded processors.
- SqueezeDet is a fully convolutional detector whose ConvDet layer computes many bounding boxes and categories before filtering them into final detections.Its pipeline first extracts a high-dimensional, low-resolution feature map with stacked convolution filters, then applies ConvDet.
2. Related Work
Related work progressed from proposal-based detectors such as R-CNN and Faster R-CNN toward single-stage and fully convolutional designs that emphasize speed, compactness, or computational reuse. In autonomous-driving detection, prior methods largely prioritized accuracy, while YOLO demonstrated real-time single-stage detection and FCNs provided a grid-output formulation.
- Proposal-based detection: R-CNN substantially improved object-detection accuracy after earlier HOG+SVM and DPM methods dominated PASCAL benchmarks.R-CNN begins by identifying region proposals likely to contain objects.
- Autonomous-driving detection: Most top-ranked published KITTI methods were based on Faster R-CNN and primarily pursued better accuracy through shallower networks or improved region proposals.The passage states that no previous methods had reported real-time inference speed, though the sentence is truncated before specifying the full claim.
- Single-stage detection: YOLO integrated region proposal and classification into one stage, making its detection pipeline extremely fast and the first CNN-based general-purpose detector to achieve real-time speed.Standard camera operation is 30 FPS, regarded as the benchmark for real-time speed.
- Compact CNN architectures: Given equal accuracy, smaller CNNs with fewer model parameters are often beneficial, motivating compact architectures for computer vision tasks.AlexNet and VGG-19 were designed for image classification and later modified for other computer vision tasks.
- Fully-convolutional networks: Fully convolutional networks produce a grid rather than a vector at the final parameterized layer, a formulation popularized for semantic segmentation and applied to other vision tasks.For image classification, convolutional grid outputs can be downsampled with average pooling instead of using fully connected layers.
3. Method Description
SqueezeDet uses a single-stage, fully convolutional detection pipeline in which ConvDet directly predicts anchored bounding boxes, confidence scores, and class probabilities. This design enables dense proposals with substantially fewer parameters than YOLO’s fully connected detection layers, followed by minimal post-processing.
- Detection Pipeline: A single network extracts a feature map, while ConvDet predicts detections over W × H uniformly distributed grid centers in one forward pass.Each bounding box carries a confidence score and C conditional class probabilities; top-N selection and NMS produce final detections.
- ConvDet Layer: ConvDet enables SqueezeDet to generate tens-of-thousands of region proposals with much fewer model parameters compared to YOLO.Unlike Faster R-CNN’s RPN, ConvDet performs detection, localization, and classification simultaneously.
- ConvDet Layer: ConvDet computes K × (4 + 1 + C) outputs at each feature-map position using K anchors with pre-selected shapes.The outputs encode four relative box coordinates, one confidence score, and C conditional class probabilities.
- Parameter Efficiency: 460X smaller: a 3x3 ConvDet layer requires approximately 0.46×10^6 parameters versus approximately 212 × 10^6 for YOLO’s two fully connected detection layers.The comparison keeps feature-map sizes, output grid centers, classes, and anchors the same: 7x7x1024, Ffc1 = 4096, K = 2, C = 20, and Wo = Ho = 7.
- Training: SqueezeDet can be trained end-to-end, similarly to YOLO, rather than using Faster R-CNN’s 4-step alternating training strategy.Its multi-task loss trains ConvDet for detection, localization, and classification, assigning ground-truth boxes to anchors with the largest overlap.
4. Experiments
Experiments on KITTI evaluated SqueezeDet and baselines across detection accuracy, recall, inference speed, model size, and resource tradeoffs. SqueezeDet variants combined competitive accuracy with high recall, real-time inference, and substantially smaller models.
- Average Precision: SqueezeDet+ achieved the highest mean average precision across all classes and difficulty levels, while matching Faster-RCNN + VGG16 on car detection accuracy.Average precision was measured on the validation set under the same training-validation split as the baselines.
- Recall: More than 80% overall recall was achieved using only the top 64 predictions; using all boxes, SqueezeDet and SqueezeDet+ reached 91% and 92%.With 1.5X image resolution, all-box recall increased to 95%.
- Speed: 19x faster inference was achieved by SqueezeDet+ than the baseline while maintaining almost the same accuracy as Faster-RCNN+VGG16.The smaller SqueezeDet reached 57.2 frames per second on a TITAN X GPU with batch size 1.
- Model size: 61X smaller model size was achieved by SqueezeDet versus Faster R-CNN + VGG16, and 30X smaller size versus Faster R-CNN + AlexNet.The comparison examined model size alongside mean average precision for car detection.
- Design space exploration: Scaling up input images decreased mAP while increasing FLOPs, reducing speed, and enlarging memory use; increasing anchors modestly raised ConvDet models’ costs.A more powerful backbone significantly improved accuracy but substantially increased FLOPs, model size, and memory footprint.
5. Conclusion
The paper proposes SqueezeDet as a fully convolutional real-time object detector designed for autonomous driving, integrating region proposal and classification while targeting compactness, speed, energy efficiency, and accuracy.
- SqueezeDet is a fully convolutional neural network for real-time object detection.The conclusion presents SqueezeDet as the paper’s central proposed model.
- ConvDet integrates region proposal and classification into a convolutional detection component that is orders of magnitude smaller than its fully connected counterpart.
- SqueezeDet and SqueezeDet+ are designed for autonomous driving with small size, fast inference, energy efficiency, and accuracy.
- 30.4x smaller model size was achieved at the same accuracy as previous baselines.
Supplementary Material: Designing Low Power Neural Network Architectures
This supplementary section is attributed to Bichen Wu, Forrest Iandola, Peter H. Jin, and Kurt Keutzer.
- The supplementary material is authored by Bichen Wu, Forrest Iandola, Peter H. Jin, and Kurt Keutzer.
1. Low Power Neural Net Design Guideline
The design guideline focuses on reducing memory accesses because off-chip DRAM is far more energy-intensive than on-chip SRAM. Fewer parameters and effective scheduling can keep consecutive-layer data in SRAM, avoiding DRAM accesses and saving energy.
- Energy-aware memory design: Two orders of magnitude more energy is consumed by a DRAM access than by an SRAM access or floating-point arithmetic operation.The work therefore focuses primarily on reducing memory accesses.
- Energy-aware memory design: Two orders of magnitude less energy is consumed by on-chip SRAM than by off-chip DRAM, while SRAM also offers lower latency and higher bandwidth.SRAM requires more transistors to store the same amount of data, so modern processors typically use substantial off-chip DRAM.
- Energy-aware memory design: Fewer model parameters and perfect scheduling can keep the parameters and activations of any two consecutive layers in on-chip SRAM, eliminating off-chip DRAM accesses.This arrangement can lead to significant energy savings.
2. Memory Footprint
SqueezeDet has a small parameter footprint that can fit on-chip SRAM, while its low activation memory and reduced FLOPs support improved energy efficiency. Its largest intermediate activation occurs at conv1.
- Parameter memory: 7.9MB is SqueezeDet’s uncompressed parameter size, enabling many processors to fit the entire model in on-chip SRAM.The model parameters can be reused during evaluations.
- Activation memory: 28.3MB is the size of SqueezeDet’s largest intermediate activation, produced by the conv1 layer.conv1 is immediately followed by a max pooling layer.
- Model comparison: SqueezeDet has a much lower activation memory footprint and performs fewer FLOPs than the compared models, supporting better energy efficiency.The comparison includes SqueezeDet variations and other models.
3. Experiments
Experiments on KITTI evaluate SqueezeDet’s energy use against prior detectors, showing substantial efficiency gains and a speed–accuracy–energy tradeoff when reducing input resolution. Comparisons with YOLO further indicate comparable energy per frame despite different input pixel counts.
- Experimental setup: Experiments evaluated energy consumption on 3,741 KITTI images at 1242x375 resolution with batch size 1 using nvidia-smi power measurements.Power readings were sampled every 0.1 second, and energy per image was computed from average working-period power and inference speed.
- Energy efficiency: 1.4J per image: SqueezeDet consumes 84× less energy than Faster R-CNN + VGG16.The comparison uses the reported energy consumption during object-detection evaluation.
- Resolution tradeoff: 1.6× faster and less than 1J per image: scaling image resolution down by 0.75× reduces mAP by 3 percentage points.This result describes the tradeoff observed for SqueezeDet under reduced input resolution.
- Comparison with YOLO: 4.2J per frame: YOLO’s reported 45FPS energy estimate is comparable with SqueezeDet+, whose input contains 2X more pixels.A direct GPU measurement gave YOLO 7.3J per frame, while the 4.2J figure uses the reported frame rate.
- Conclusion: The experiments show that SqueezeDet and its variations are very energy efficient compared with previous neural network based object detectors.This is the paper’s overall experimental conclusion.