Scribd
Upload
206 views

Object Detection and Its Implementation On Android Devices

The document summarizes research on implementing an object detection algorithm on Android devices. Key points: 1. The researchers implemented a CNN-based object detection model called SqueezeDet, which is based on SqueezeNet, on Android for tasks like autonomous driving and surveillance. 2. The model achieves 81% recall on the KITTI dataset with low computation cost and memory demand of around 8MB, making it suitable for mobile devices. 3. Testing showed the model is most sensitive to blurry images, especially for cyclist classification, but is otherwise robust to other image variations.

Uploaded by

api-376298381
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
206 views

Object Detection and Its Implementation On Android Devices

The document summarizes research on implementing an object detection algorithm on Android devices. Key points: 1. The researchers implemented a CNN-based object detection model called SqueezeDet, which is based on SqueezeNet, on Android for tasks like autonomous driving and surveillance. 2. The model achieves 81% recall on the KITTI dataset with low computation cost and memory demand of around 8MB, making it suitable for mobile devices. 3. Testing showed the model is most sensitive to blurry images, especially for cyclist classification, but is otherwise robust to other image variations.

Uploaded by

api-376298381
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

Object Detection and Its Implementation on Android Devices

Rao Zhang, Zhongjie Li


(zhangrao/jay2015)@stanford.edu

Introduction Approach Results


Object detection is a very important task for different applications 1. Training Results
including autonomous driving, face detection, video surveillance, etc. CNN 1. Android App Architecture
The squeezeDet model is trained with KITTI detection dataset. After ~35k steps of training,
based algorithm could be a great solution for object detection with high the overall recall can get 81%. The detection precisions are as following.
accuracy. Besides, most current deep learning applications are running on detection accuracy car cyclist pedestrian
servers or desktop computers. Considering there are a lot of mobile easy 90% 86% 80%
computing devices available, we implemented the CNN based object
medium 85% 80% 74%
detection algorithm on Android devices. The model architecture is based on
hard 75% 77% 67%
SqueezeNet and further improved to find bounding boxes for recognized
objects. The total model size is around 8 MB and makes it run fast,
especially on mobile devices.

Problem Statement 2. SqueezeNet + ConvNet Graph[1]


To find and implement an object detection algorithm thats suitable for
implementation on mobile platforms. The key requirements include:
Low computation cost 2. Model Sensitivity to Image Variations
Low memory demand The model is validated against multiple image viations, including blur, brightness, contrast,
High efficiency color and resolution. The results showed that the model is most sensitive to blur, especially
for cyclist classifications.

Datasets and Preprocessing


Data Sets
The dataset we use is the object detection data-set of KITTI Vision
Benchmark Suite, which contains 7481 training images and 7518 test
images. Total 80256 objects are labeled for this dataset and the 3 classes
used for evaluation are cars, pedestrians and cyclists. The distribution of
object number in the training data-set is shown below. 51865 objects are
labeled, including 28742 cars, 4487 pedestrians and 1627 cyclists. On
average: 3.8 cars, 0.6 pedestrian and 0.2 cyclist per image.
Data processing
Image flipping
Random cropping
Batch normalization
Image variations of brightness/contrast/blur, etc.

[1] Wu, Bichen, et al. "SqueezeDet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving." arXiv preprint arXiv:1612.01051 (2016).
[2] Iandola, Forrest N., et al. "SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size." arXiv preprint arXiv:1602.07360 (2016)

You might also like