cvpaper.challenge in 2015 - A review of CVPR2015 and DeepSurvey

Hirokatsu Kataoka, Yudai Miyashita, Tomoaki K. Yamabe, Soma Shirakabe, Shin-ichi Sato, Hironori Hoshino, Ryo Kato, Kaori Abe, Takaaki Imanari, Naomichi Kobayashi, Shinichiro Morita, Akio NakamuraarXiv 2016 · OpenReview

Abstract

The rapid expansion of the computer vision field presents a significant challenge for researchers attempting to maintain a comprehensive understanding of state-of-the-art developments. The cvpaper.challenge group addressed this by systematically reviewing all 602 papers accepted to CVPR 2015. This initiative not only identified key research trends—most notably the pervasive adoption of deep learning—but also introduced "DeepSurvey," a novel collaborative research framework modeled after neural network architectures. This paper details the findings of the CVPR 2015 survey and the conceptual foundations of the DeepSurvey methodology.

1. Introduction

The volume of research output in computer vision (CV) has reached a scale where individual researchers struggle to synthesize the "big picture." To mitigate this, the cvpaper.challenge group was formed to collaboratively read, summarize, and categorize major conference proceedings. This paper reviews their analysis of CVPR 2015, a landmark year that saw deep learning transition from a specialized tool to a dominant paradigm across nearly all sub-domains of vision.

2. Methodology: The DeepSurvey Framework

A core contribution of the cvpaper.challenge is the "DeepSurvey" methodology. Recognizing that research is a process of information processing, the authors proposed a workflow that mimics a multi-layer Convolutional Neural Network (CNN):

  • Input Layer: Systematic knowledge acquisition through the reading of international conference papers.
  • Convolution/Pooling Layers: Iterative cycles of individual idea generation followed by group discussions to filter and consolidate insights.
  • Fully Connected Layer: The transition from theoretical understanding to practical implementation via hackathons and experimental validation.
  • Output: The generation of a new, high-quality research paper.

This "Human Neural Network" approach leverages collective intelligence to process information more effectively than any single researcher could achieve in isolation.

The survey categorized the 602 papers of CVPR 2015 into three primary domains: Recognition, 3D, and Imaging/Image Processing. The rigorous ~20% acceptance rate of the conference ensured that the surveyed papers represented the highest quality of contemporary research.

3.1 The Dominance of Deep Learning

Deep learning (DL) was present in approximately 40% (250 of 602) of the accepted papers. While its roots were in classification—with GoogLeNet serving as the benchmark to beat following its ILSVRC2014 win—2015 marked its aggressive expansion into 3D feature extraction, stereo matching, and image restoration tasks like deblurring.

3.2 From Classification to Semantic Understanding

A significant shift was observed from simple object recognition toward complex scene understanding. Key developments included:

  • Semantic Segmentation: The rise of Fully Convolutional Networks (FCNs) for pixel-wise labeling.
  • Vision and Language: The emergence of image captioning and visual question answering, bridging the gap between CV and Natural Language Processing (NLP).

3.3 Real-time 3D and Large-scale Vision

The Best Paper Award for DynamicFusion underscored a trend toward real-time, non-rigid 3D reconstruction using consumer sensors. Simultaneously, the field moved toward "world-scale" problems, exemplified by Structure from Motion (SfM) applications involving 100 million images from Flickr.

4. Key Datasets

The 2015 cycle highlighted several benchmarks that have since become foundational:

  • SUN RGB-D: A large-scale indoor 3D dataset containing over 10,000 RGB-D images.
  • ActivityNet: A benchmark for video-based human activity understanding, scaling vision tasks to the temporal domain.
  • Microsoft COCO: A catalyst for advancements in object detection, segmentation, and captioning.

5. Conclusion

The cvpaper.challenge 2015 initiative demonstrated that systematic, collaborative literature review is essential for navigating the modern CV landscape. By modeling the research process as a neural network, the DeepSurvey framework provides a scalable way to transform the vast "input" of published research into meaningful "output" in the form of new ideas and implementations.

Summary updated

Abstract

The "cvpaper.challenge" is a group composed of members from AIST, Tokyo Denki Univ. (TDU), and Univ. of Tsukuba that aims to systematically summarize papers on computer vision, pattern recognition, and related fields. For this particular review, we focused on reading the ALL 602 conference papers presented at the CVPR2015, the premier annual computer vision event held in June 2015, in order to grasp the trends in the field. Further, we are proposing "DeepSurvey" as a mechanism embodying the entire process from the reading through all the papers, the generation of ideas, and to the writing of paper.
BibTeX
@misc{kataoka2016cvpaper,
  author = {Hirokatsu Kataoka and Yudai Miyashita and Tomoaki K. Yamabe and Soma Shirakabe and Shin-ichi Sato and Hironori Hoshino and Ryo Kato and Kaori Abe and Takaaki Imanari and Naomichi Kobayashi and Shinichiro Morita and Akio Nakamura},
  title = {{cvpaper.challenge in 2015 - A review of CVPR2015 and DeepSurvey}},
  year = {2016},
  eprint = {1605.08247},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1605.08247},
}
Figure