计算机视觉这一年（英文版）_54页_2mb

2016 in Computer Vision: A Year of Breakthroughs

Core Content Overview

This document provides a summary of the most significant advancements in Computer Vision in 2016, with a focus on classification/localisation, object detection, object tracking, segmentation, super-resolution, style transfer, and colourisation. It highlights the evolution of deep learning techniques and their impact on both research and real-world applications.

Main Topics and Key Insights

Classification/Localisation

• Definition: Involves assigning a label to an image and identifying the location of objects within it.
• Advancements:
  • Trimps-Soushen won with a 2.99% top-5 classification error and 7.71% localisation error.
  • ResNeXt achieved 3.03% top-5 classification error, improving upon ResNet.
  • YOLO9000 introduced joint training for detection and classification, allowing real-time detection across 9000+ categories.
  • SSD achieved 75.1% mAP, outperforming Faster R-CNN in speed and efficiency.
• Trends:
  • Shift towards end-to-end training for improved efficiency.
  • Feature Pyramid Networks (FPN) enabled multi-scale detection without sacrificing speed or memory.

Object Detection

• Definition: Detects multiple objects in an image, providing bounding boxes and labels.
• Key Systems:
  • YOLOv2 improved upon YOLO with better accuracy and speed.
  • R-FCN achieved 170ms per image and outperformed Faster R-CNN in speed.
  • Faster R-CNN remained highly effective, especially when combined with advanced feature extractors.
• Results:
  • COCO 2016 Detection Challenge: Google's G-RMI achieved 41.5% AP, a 4.2% increase from the previous year.
  • ImageNet LSVRC: CUImage achieved 66% meanAP, winning 109 out of 200 object categories.

Object Tracking

• Definition: Follows objects across video sequences.
• Notable Methods:
  • Fully-Convolutional Siamese Networks achieved SOTA and operated at real-time frame rates.
  • Deep Motion Features were introduced for visual tracking, combining hand-crafted and deep features.
  • Virtual Worlds as Proxy created synthetic environments with full labels to improve benchmark diversity.
  • GOTURN achieved 100 FPS tracking with deep regression networks.
• Key Contributions:
  • Addressed issues like object occlusion and appearance variation.
  • Deep Motion Features were used for the first time in visual tracking and won the Best Paper at ICPR 2016.

Segmentation

• Definition: Divides images into pixel-level groupings and labels them.
• Key Techniques:
  • DeepMask and SharpMask were introduced by FAIR, with SharpMask refining segmentation details.
  • MultiPathNet identified objects based on masks.
  • DeepLab achieved promising results in semantic segmentation.
  • Weakly Supervised Learning was used to reduce the need for full supervision.
• Applications:
  • Healthcare: Segmentation was applied to colonoscopy, MRI, 3D ultrasound, and retinal vessel images.
  • Connectomics: FusionNet was benchmarked against SOTA EM segmentation methods.

Super-resolution, Style Transfer & Colourisation

• Super-resolution:
  • RAISR from Google offered a fast and memory-efficient method for image enhancement.
  • SRGAN and SRResNet were among the top methods for super-resolution, with SRGAN achieving better texture realism and Mean Opinion Score (MOS).
• Style Transfer:
  • Neural Algorithm of Artistic Style (2015) was expanded upon in 2016.
  • Caffe2Go enabled mobile integration of style transfer.
  • Conditional Instance Normalisation allowed a single network to capture 32 styles simultaneously.
• Colourisation:
  • Automated methods were developed to colorize monochrome images in a way that mimics human perception.
  • These techniques use real-world knowledge to ensure color consistency with the image's context.

Key Information

• CNNs (Convolutional Neural Networks) dominated 2016, with AlexNet (2012) sparking a revolution in the field.
• End-to-end training became a major trend, improving efficiency and accuracy.
• Open-source tools and public datasets (e.g., ImageNet, COCO) were critical in driving progress.
• Real-world applications included healthcare, autonomous driving, and content creation.

Conclusion

2016 marked a significant year in Computer Vision, with breakthroughs in deep learning architectures, efficient training methods, and practical applications. The field continues to evolve rapidly, driven by both academic research and commercial interest, with a growing emphasis on real-time performance and multi-domain adaptability.