PaperlayerView exact revision on arXiv

Source-linked AI summary

YOLO-LITE: A Real-Time Object Detection Algorithm Optimized for Non-GPU Computers

Jonathan Pedoeem, Rachel Huang

arXiv:1811.05588v1cs.CV

TL;DR

Object detection systems need real-time performance on computationally limited devices, but existing approaches may be too slow without GPUs. The paper develops YOLO-LITE from YOLO using a lightweight architecture and evaluates its speed against other fast detectors. YOLO-LITE reaches 21 FPS on a non-GPU computer, while remaining substantially faster than SSD and Tiny-YOLOV2.

  • Problem

    Real-time object detection remains difficult on non-GPU devices because computationally heavy models can be too slow for practical use.

  • Method

    YOLO-LITE adapts YOLO into a shallow lightweight network and removes batch normalization to reduce feedforward computation.

  • Results

    21 FPS on a non-GPU computer, with YOLO-LITE reported as 3.6× faster than SSD and 8.8× faster than Tiny-YOLOV2.

  • Takeaways & Limitations

    YOLO-LITE demonstrates that shallow networks can support lightweight real-time object detection and that batch normalization may warrant reconsideration in smaller networks.

  • Takeaways & Limitations

    YOLO-LITE's accuracy needs improvement for the model to be viable in applications such as autonomous vehicles.

Abstract

from arXiv · show

This paper focuses on YOLO-LITE, a real-time object detection model developed to run on portable devices such as a laptop or cellphone lacking a Graphics Processing Unit (GPU). The model was first trained on the PASCAL VOC dataset then on the COCO dataset, achieving a mAP of 33.81% and 12.26% respectively. YOLO-LITE runs at about 21 FPS on a non-GPU computer and 10 FPS after implemented onto a website with only 7 layers and 482 million FLOPS. This speed is 3.8x faster than the fastest state of art model, SSD MobilenetvI. Based on the original object detection algorithm YOLOV2, YOLO- LITE was designed to create a smaller, faster, and more efficient model increasing the accessibility of real-time object detection to a variety of devices.

I. INTRODUCTION

Real-time object detection must balance accuracy with computational limits, especially on devices without GPUs. YOLO-LITE addresses this gap by pursuing real-time detection on standard non-GPU computers.

  • Real-time detection is necessary for autonomous vehicles, augmented reality devices, and other intelligent systems operating in changing environments.
  • Larger and deeper networks can improve accuracy but may impose computational costs that hinder real-time performance on limited platforms.
  • YOLO-LITE uses YOLO as its starting point to target real-time object detection on a standard non-GPU computer.
  • CNNs use convolutional filters to create outputs from input regions while learning fewer weights than fully connected layers.

A. R-CNN

YOLO combines object localization and classification in one process, using grid-based bounding-box confidence and class predictions. Its accuracy is evaluated with average precision, while earlier YOLO variants remained slow on non-GPU computers.

  • YOLO: YOLO performs bounding-box and class predictions simultaneously after evaluating the input image once.The input image is divided into an S × S grid, with B bounding boxes and confidence scores in each cell.
  • YOLO: Intersection over union (IOU) measures overlap between predicted and ground-truth boxes, with values closer to 1 indicating closer alignment.IOU ranges from 0 to 1 and compares overlapping area with total union area.
  • YOLO: Class-specific probabilities are predicted for each grid cell alongside the bounding boxes.The class-specific term combines conditional class probability, object probability, and IOU.
  • YOLO: The loss function corrects box centers, dimensions, confidence, and classification predictions during optimization.Its weighting emphasizes boxes containing objects and reduces emphasis on boxes without objects; lower loss indicates higher performance.
  • YOLO: Average precision summarizes detection accuracy from precision at threshold k and changes in recall.The paper defines P(k) as precision at threshold k and Δr(k) as the change in recall.
  • YOLO: Tiny-YOLOv2 achieves 23.7% mAP and 5.41 billion FLOPS, but its speed drops from 244 FPS to about 2.4 FPS on a non-GPU laptop.This reduction limits real-time object detection access on many laptops and cellphones without GPUs.

III. YOLO-LITE ARCHITECTURE

YOLO-LITE targets real-time object detection on non-GPU computers through a shallow architecture and iterative modification of Tiny-YOLOv2. Its design prioritizes speed while examining the roles of network depth, FLOPS, and batch normalization.

  • III. YOLO-LITE ARCHITECTURE: YOLO-LITE targets at least ∼10 FPS on a non-GPU computer with 30% mAP on PASCAL VOC.This benchmark defines the architecture’s stated performance goal.
  • III. YOLO-LITE ARCHITECTURE: YOLO-LITE demonstrates shallow networks for fast non-GPU object detection applications.The approach uses existing components rather than introducing an original convolution layer or pruning method.
  • III. YOLO-LITE ARCHITECTURE: The paper suggests batch normalization is unnecessary for shallow networks and slows overall network speed.This claim is presented as one of YOLO-LITE’s main contributions.
  • III. YOLO-LITE ARCHITECTURE: Using Tiny-YOLOv2 as a starting point, the authors iteratively removed and added layers and trained models on PASCAL VOC 2007 and 2012.All iterations retained Tiny-YOLOv2’s final layer, and validation-set mAP guided subsequent iterations.
  • III. YOLO-LITE ARCHITECTURE: FLOPS correlated with FPS, whereas adding filters or layers did not reliably improve mAP.The authors report difficulty identifying indicators that would predict improvements in both metrics.

A. Setup

The models were developed and tested with Darknet-based tooling, using a GPU-equipped training computer and a Dell XPS 13 laptop for FPS testing.

  • A. Setup: Darknet was used to train and test the models, while FPS testing used Darkflow’s live demo script.Training ran on an Alienware Aura R7 with an Intel i7 CPU and Nvidia 1070 GPU; FPS testing ran on a Dell XPS 13 laptop.

B. PASCAL VOC and COCO Datasets

YOLO-LITE was evaluated on PASCAL VOC and then retrained on COCO, spanning datasets with different class counts and training-set sizes.

  • B. PASCAL VOC and COCO Datasets: COCO 2014 contains 80 classes and approximately 40,000 training images.The highest-performing PASCAL VOC model was retrained on this second dataset.

C. Indicators for Speed and Precision

The experiments assessed which architectural and training indicators tracked YOLO-LITE’s speed and precision. FLOPS and epoch rate related to speed, while added layers and filters did not reliably improve accuracy, and smaller input images traded mAP for faster inference.

  • Indicators for Speed and Precision: FLOPS count and epoch rate were useful indicators of FPS, although epoch rate was more informative because FLOPS excludes batch-normalization computation.Loss was not an exact indicator of mAP because the listed losses combine features rather than directly defining mAP.
  • Indicators for Speed and Precision: Trials 4, 5, 8, and 10 showed no clear relationship between adding layers and filters and improving accuracy.
  • Image Size: 6.94 FPS vs 2.4 FPS resulted when halving input image size, while mAP changed from 40.48% to 30.24%.The authors preferred the speed increase over the mAP decrease for their purposes.

E. Batch Normalization

The paper examines batch normalization as a training aid and finds that YOLO-LITE’s shallow architecture does not require it. The computation between layers was considered a source of feedforward slowdown.

  • Batch Normalization: Batch normalization transforms one layer’s output to mean zero, variance one, and standard deviation one before the next layer.It addresses differing mini-batch distributions, which can contribute to covariate shift and vanishing gradients during training.
  • Batch Normalization: YOLO-LITE’s small network does not suffer greatly from covariate shift or vanishing gradients, so the paper argues batch normalization is unnecessary.
  • Batch Normalization: Batch-normalization calculations between layers were considered to hold up the network and slow feedforward processing.Each input value must be updated during feedforward computation.
  • Pruning: Pruning, quantization, and Huffman coding had been suggested to shrink networks and speed them up, but pruning YOLO-LITE produced no improvement in accuracy or speed.The paper attributes this difference to YOLO-LITE consisting mainly of convolutional rather than fully connected layers.

IV. RESULTS

YOLO-LITE was evaluated through 18 trials developed from Tiny-YOLOv2 and the best trial was additionally tested on COCO. Its best architecture used substantially fewer layers, filters, and FLOPS than Tiny-YOLOv2.

  • Results: 18 trials were attempted, with Figure 4 reporting mAP and FPS for each trial and Tiny-YOLOv2.
  • Results: The best trial developed on PASCAL VOC was also run on COCO, with the top results reported for both datasets.
  • Architecture: 9 convolutional layers, 3,181 filters, and 6.97 billion FLOPS describe Tiny-YOLOv2.
  • Architecture: 7 layers, 749 filters, and 482 FLOPS describe YOLO-LITE Trial 3-no batch, which used 14× fewer FLOPS than Tiny-YOLOv2.The paper links the lighter model to YOLO-LITE’s faster performance.

V. COMPARISON WITH OTHER FAST OBJECT DETECTION NETWORKS

The comparison places YOLO-LITE among fast object-detection networks for non-GPU devices. It reports faster performance than SSD and Tiny-YOLOv2, with lower website FPS than local laptop execution.

  • Comparison: ∼2.4 FPS is reported for YOLO’s tiny architecture on non-GPU computers, making it hardly real-time in that setting.
  • Comparison: 5.8 FPS and 21% mAP are reported for SSD Mobilenet V1 on a non-GPU laptop.
  • Comparison: 3.6× faster than SSD and 8.8× faster than Tiny-YOLOv2 describes YOLO-LITE’s comparison in Table VIII.
  • Web Deployment: 21 FPS locally on a Dell XPS 13 laptop decreased to around 10 FPS after YOLO-LITE was deployed on the website.The reported FPS may differ depending on the device.

VI. CONCLUSION

YOLO-LITE brings object detection to non-GPU computers through a shallow, lightweight design, while exposing a substantial speed–accuracy tradeoff. The paper identifies several future directions for improving accuracy, efficiency, and model size.

  • YOLO-LITE brings real-time object detection to non-GPU computers and runs at 21 FPS on such a computer.
  • YOLO-LITE shows that shallow networks have potential for lightweight real-time object detection and questions batch normalization in smaller shallow networks.
  • 12.16% mAP marks a substantial accuracy decrease from 48.1% for YOLOv2 and 23.7% for YOLOv2-Tiny.
  • The paper identifies accuracy as the main barrier to YOLO-LITE’s viability in applications such as autonomous vehicles.
  • Future improvements include multiple prediction locations, filter pruning, and ShuffleNet-style group convolution with channel shuffling.

VIII. RELEVANT LINKS

The paper provides online resources for the YOLO-LITE web implementation and training files for PASCAL VOC and COCO.

  • The web implementation is available at reu2018dl.github.io, while configuration and weights files are available on the yolo-lite GitHub repository.
Loading 1811.05588v1…