计算机视觉这一年（英文版）_56页-1mb

2016 Year in Computer Vision Summary

Core Content Overview

This document provides an overview of the major advancements in Computer Vision in 2016, focusing on classification/localisation, object detection, object tracking, segmentation, super-resolution, style transfer, colourisation, and action recognition. It also touches upon ConvNet architectures, datasets, and trends in the field. The document is structured into four main parts and includes references to academic publications, datasets, and notable projects.

---

Part One: Classification/Localisation, Object Detection, Object Tracking

Classification/Localisation

• Definition: Assigning a label to an entire image (e.g., 'cat') and identifying the location of the object using bounding boxes.
• Advancements:
  • Techniques have advanced to the point where they surpass human performance on ImageNet.
  • New datasets like Places2 with 8 million images and 365 scene categories were introduced.
• Notable Results:
  • Hikvision won with a 9% top-5 error using an ensemble of deep Inception-style networks.
  • Trimps-Soushen achieved 2.99% top-5 classification error and 7.71% localisation error.
  • ResNeXt came close with 3.03% top-5 classification error using an extended ResNet architecture.

Object Detection

• Definition: Detecting multiple objects in an image and outputting their bounding boxes and labels.
• Key Trends:
  • Shift towards end-to-end training for faster and more efficient detection systems.
  • Techniques like YOLOv2, SSD, and R-FCN were highlighted for their speed and accuracy.
• Notable Systems:
  • YOLO9000: Achieved 41.5% AP in the COCO 2016 Detection Challenge.
  • SSD: Achieved 75.1% mAP, outperforming Faster R-CNN.
  • R-FCN: Achieved 170ms per image with improved accuracy over traditional methods.
• Comparative Results:
  • ImageNet LSVRC:
    • DET: 66% meanAP by CUImage.
    • VID: 80.8% meanAP by NUIST.
    • CUvideo: 55.8% meanAP for video with tracking.

Object Tracking

• Definition: Following specific objects in a video sequence.
• Key Papers:
  • Fully-Convolutional Siamese Networks: Achieved SOTA and real-time performance.
  • Deep Motion Features: Fused hand-crafted, RGB, and motion features for improved tracking.
  • Virtual Worlds as Proxy: Created synthetic environments for multi-object tracking with full labels.
  • Globally Optimal Object Tracking: Used Dynamic Programming for handling occlusion.
• Applications: Crucial in autonomous driving, robotics, and video analysis.

---

Part Two: Segmentation, Super-res/Colourisation/Style Transfer, Action Recognition

Segmentation

• Definition: Dividing images into pixel-level groupings for classification and labelling.
• Types:
  • Semantic Segmentation: Assigns labels to each pixel.
  • Instance Segmentation: Differentiates between instances of the same class.
• Key Advancements:
  • DeepLab achieved encouraging results in semantic segmentation.
  • SharpMask improved on DeepMask by refining object masks.
  • MultiPathNet identified objects delineated by masks.
• Healthcare Applications:
  • Colonoscopy Image Segmentation.
  • MRI Subcortical Segmentation.
  • Brain Lesion Detection using semi-supervised learning.
  • Retinal Vessel Segmentation and Glioblastoma Segmentation using 3D CNNs.
  • FusionNet for connectomics benchmarking.

Super-resolution, Style Transfer & Colourisation

• Super-resolution: Enhancing low-resolution images to high-resolution.
• Key Techniques:
  • RAISR: A fast and memory-efficient method using filter training.
  • SRGAN: Uses GANs to generate photo-realistic textures with 4x upscaling.
  • SRResNet: Best in PSNR but not in MOS (Mean Opinion Score).
• Applications:
  • Mobile devices: Facebook's Caffe2Go enabled super-resolution on phones.
  • Image enhancement: Inspired by "CSI Miami" style image processing.

Style Transfer & Colourisation

• Style Transfer: Applying the artistic style of one image to another.
• Key Papers:
  • Nikulin & Novak (2016): Transferred style to a cat image.
  • Dumoulin et al. (2017): Enabled 32 styles to be captured simultaneously.
• Applications:
  • Game asset creation.
  • Mobile integration via Facebook and Google.
  • Artistic applications through tools like Prisma and Artomatix.

---

Part Three: Toward a 3D Understanding of the World

• This section discusses 3D understanding in Computer Vision, though it is not explicitly detailed in the text.
• It highlights the complexity and significance of true visual understanding and its role in Strong AI development.
• The field is still in early stages, but 2016's advancements laid the groundwork for future progress.

---

Part Four: ConvNet Architectures, Datasets, Ungroupable Extras

ConvNet Architectures

• CNNs (Convolutional Neural Networks) remain the workhorse of Computer Vision.
• AlexNet in 2012 sparked a revolution in the field.
• ResNet, ResNeXt, Inception, and Xception were widely used in 2016.
• End-to-end training is emphasized as a way to reduce computational overhead.

Datasets

• ImageNet and ILSVRC were central to object detection and classification.
• COCO (Common Objects in Context) is a smaller, more curated dataset for scene understanding.
• Places2 was used for scene classification with 365 categories.

Ungroupable Extras & Interesting Trends

• SqueezeDet and PVANet demonstrated cheap and efficient object detection.
• Video Propagation Networks enabled mask propagation through video sequences.
• 3D CNNs were used in medical imaging for segmentation.
• GANs became a key tool for super-resolution and style transfer.

---

Conclusion

• Computer Vision in 2016 saw significant progress in classification, detection, tracking, and segmentation.
• End-to-end learning and CNNs were central to these advancements.
• Datasets like ImageNet, COCO, and Places2 played a crucial role in benchmarking and research.
• Super-resolution and style transfer expanded the practical applications of Computer Vision beyond traditional tasks.
• Future potential: The field is seen as a stepping stone toward Strong AI, with 3D understanding and contextual awareness being the next frontiers.
• Open source and collaboration were key enablers of innovation, with deep learning frameworks and publicly available resources facilitating broader access and experimentation.