[Paper Review] Fast R-CNN

Ross Girshick|arXiv (Cornell University)|Apr 30, 2015

Advanced Neural Network Applications24 references1,768 citations

TL;DR

Fast R-CNN proposes an efficient, accurate region-based convolutional network for object detection that accelerates training and inference by sharing convolutional feature computation across regions. It achieves 9x faster training and 213x faster testing than R-CNN on PASCAL VOC 2012, while improving mean average precision (mAP) using the VGG16 network.

ABSTRACT

This paper proposes a Fast Region-based Convolutional Network method (Fast R-CNN) for object detection. Fast R-CNN builds on previous work to efficiently classify object proposals using deep convolutional networks. Compared to previous work, Fast R-CNN employs several innovations to improve training and testing speed while also increasing detection accuracy. Fast R-CNN trains the very deep VGG16 network 9x faster than R-CNN, is 213x faster at test-time, and achieves a higher mAP on PASCAL VOC 2012. Compared to SPPnet, Fast R-CNN trains VGG16 3x faster, tests 10x faster, and is more accurate. Fast R-CNN is implemented in Python and C++ (using Caffe) and is available under the open-source MIT License at this https URL.

Motivation & Objective

• To address the inefficiency of R-CNN in training and inference by reducing redundant computation across region proposals.
• To improve detection accuracy while significantly reducing training and testing time compared to prior methods like R-CNN and SPPnet.
• To enable end-to-end training of deep networks with shared convolutional features across regions.
• To develop a scalable, fast, and accurate object detection framework suitable for real-world deployment.

Proposed method

• It uses a shared convolutional neural network to extract features from the full image once, rather than separately for each region proposal.
• It applies a RoI (Region of Interest) pooling layer to extract fixed-size features from each region proposal using the shared convolutional features.
• It replaces the selective search method with a region proposal network (RPN) in later versions, though this paper focuses on the original Fast R-CNN with precomputed proposals.
• It enables joint training of the R-CNN head and the region proposal network via a multi-task loss function combining classification and bounding box regression.
• It uses a softmax classifier with a softmax loss and a bounding box regression head with a smooth L1 loss for end-to-end optimization.
• The framework is implemented in Python and C++ using the Caffe deep learning framework, enabling high-speed inference and training.

Experimental results

Research questions

• RQ1 Can region-based object detection be made significantly faster without sacrificing accuracy?
• RQ2 How can shared convolutional feature computation improve training and inference efficiency in R-CNN-style detectors?
• RQ3 Can end-to-end training of the full detection pipeline be achieved with high accuracy and speed?
• RQ4 How does Fast R-CNN compare to R-CNN and SPPnet in terms of training speed, inference speed, and mAP on standard benchmarks?

Key findings

• Fast R-CNN trains the VGG16 network 9 times faster than R-CNN on the PASCAL VOC 2012 dataset.
• It achieves 213 times faster inference speed than R-CNN, significantly improving real-time detection potential.
• Compared to SPPnet, Fast R-CNN trains VGG16 3 times faster and tests 10 times faster.
• Fast R-CNN achieves a higher mean average precision (mAP) than both R-CNN and SPPnet on the PASCAL VOC 2012 benchmark.
• The method enables end-to-end training with shared convolutional features, reducing redundancy and improving optimization.
• The framework is open-sourced under the MIT License and is available for public use with Caffe.