[Paper Review] ChainerCV: a Library for Deep Learning in Computer Vision

Yusuke Niitani, Toru Ogawa+2 more|arXiv (Cornell University)|Aug 28, 2017

Advanced Neural Network Applications10 references18 citations

TL;DR

ChainerCV is a deep learning library for computer vision that provides high-quality, reproducible implementations of state-of-the-art models like Faster R-CNN, SSD, and SegNet, with training code guaranteed to match published performance. Built on Chainer and CuPy, it enables flexible, GPU-accelerated training and inference with unified data handling, visualization, and evaluation tools for research and development.

ABSTRACT

Despite significant progress of deep learning in the field of computer vision, there has not been a software library that covers these methods in a unifying manner. We introduce ChainerCV, a software library that is intended to fill this gap. ChainerCV supports numerous neural network models as well as software components needed to conduct research in computer vision. These implementations emphasize simplicity, flexibility and good software engineering practices. The library is designed to perform on par with the results reported in published papers and its tools can be used as a baseline for future research in computer vision. Our implementation includes sophisticated models like Faster R-CNN and SSD, and covers tasks such as object detection and semantic segmentation.

Motivation & Objective

• To address the lack of a unified, high-quality software library that implements and reproduces deep learning-based computer vision methods with consistent performance.
• To lower the barrier to entry for researchers and developers by providing a simple, flexible, and well-documented interface for training and deploying complex vision models.
• To ensure training code achieves performance on par with original papers, enabling reproducibility and serving as a reliable baseline for future research.
• To support end-to-end workflows including data loading, training, evaluation, and visualization for tasks like object detection and semantic segmentation.
• To facilitate integration into larger systems, such as robotics pipelines, by offering modular, reusable components for vision algorithms.

Proposed method

• Leverages Chainer and CuPy to enable GPU-accelerated computation with NumPy-like syntax and automatic differentiation.
• Provides a unified data interface where images are represented as (C, H, W) arrays in RGB format, ensuring consistency across models.
• Implements training pipelines for Faster R-CNN, SSD300/512, and SegNet using standard datasets (PASCAL VOC, CamVid) and standard hyperparameters.
• Uses TransformDataset to apply data augmentation and preprocessing, including random scaling and color jittering, during training.
• Integrates evaluation metrics such as mean average precision (mAP), pixel accuracy, mean pixel accuracy, and mean IoU for object detection and segmentation tasks.
• Offers visualization tools that render predictions and ground truth on input images, enabling model analysis and debugging.

Experimental results

Research questions

• RQ1 Can a deep learning library provide training code that reproduces the performance of published state-of-the-art models in object detection and semantic segmentation?
• RQ2 How can a software library improve usability and reproducibility in computer vision research while maintaining high performance?
• RQ3 To what extent can a unified, modular library support complex vision pipelines, including data loading, training, evaluation, and visualization?
• RQ4 Can reference implementations of sophisticated models like Faster R-CNN and SSD be made accessible and reliable for researchers with limited deep learning experience?
• RQ5 How does the performance of a library's training implementation compare to the original paper's reported results across multiple benchmarks and models?

Key findings

• The ChainerCV implementation of Faster R-CNN with VGG-16 achieved a mean average precision (mAP) of 70.5% on PASCAL VOC 2007 test, matching the original paper's reported result of 69.9%.
• The SSD512 implementation achieved an mAP of 80.1%, slightly exceeding the original paper's reported result of 79.5%.
• The SSD300 implementation matched the original paper's mAP of 77.5%.
• The SegNet implementation achieved 82.8% pixel accuracy, 67.1% mean pixel accuracy, and 47.2% mean IoU on CamVid, outperforming the original paper's results of 82.7%, 62.3%, and 46.3% respectively.
• The library's training code consistently reproduced or slightly exceeded published performance across all evaluated models, demonstrating high reproducibility.
• The use of CuPy enabled efficient GPU computation with minimal setup overhead, and the library's modular design allowed seamless integration into larger software systems.