After defining AI-related research topics across all of science in Defining Computer Vision, Natural Language Processing, and Robotics Research Clusters, here we explore the 1,105 research clusters that are labeled as computer vision RCs (as of February 2021).1 RCs are assigned a CV label if they have at least 25 percent AI-related papers and 25 percent CV-related papers, with CV being the dominant AI-related topic (i.e., natural language processing and robotics have lower percentages). We investigate how artificial intelligence-related CV research is developed and applied across all of science and provide details on RCs with low and high CV-related paper concentrations.
Applied in a wide range of domains, from autonomous vehicles to medical imaging, CV technologies utilize AI/machine learning methods on visual inputs, such as images or videos. Tumor recognition in medical imaging, for instance, is an example of applied CV technology.2 Figure 1 displays CV RCs highlighted in the Map of Science, where each RC is colored by its broad area of research.
Figure 1. CV RCs Highlighted in the Map of Science.
Table 1 provides the breakdown of CV RCs by their broad area of research, with the overwhelming majority falling under computer science.
Table 1. Number of CV RCs by Broad Research Area
Broad Research Area | Number of CV RCs | Percentage of CV RCs |
Computer Science | 1,024 | 93% |
Mathematics | 24 | 2% |
Medicine | 24 | 2% |
Earth Science | 21 | 2% |
Social Science | 6 | 0.5% |
Biology | 5 | 0.5% |
Engineering | 1 | 0.1% |
Table 2 provides the breakdown of CV RCs by their concentration of CV-related publications. We find that 43 percent of CV RCs are made up of anywhere between 25 percent and 50 percent CV-related papers.
Table 2. CV-related Publication Concentrations Across CV RCs
Percentage of CV-related Publications | Number of RCs |
[25%, 50%] | 473 |
(50%, 75%] | 370 |
(75%, 100%] | 262 |
Total | 1,105 |
In order to understand the range of RCs that can be assigned the CV label, we provide details on four RCs:
- The CV RC with the highest percentage of CV-related publications
- The CV RC with the lowest percentage of CV-related publications
- A CV RC in non-computer science STEM field
- A CV RC in a non-STEM field
For each of these RCs, we provide the top five core papers. Core papers are publications that have strong citation links within an RC, meaning that they have high citation counts from the other publications in that cluster. Since RCs do not necessarily represent a homogenous area of research, we can review the member publications to identify the central areas of research that a RC is focused on.
CV-related RC with the highest percentage of CV-related publications:
With 679 papers between 2015 and 2021, RC 35520 focuses on object segmentation in video footage. More generally described, this RC focuses on pattern recognition. Papers in this RC range from data set creation to proposing new methods of object segmentation.3 98 percent of papers in RC 35520 are CV-related, as the main area of research is pattern recognition, a subset of CV research.
RC 35520 Top Five Core Papers:
- Learning Video Object Segmentation from Static Images
- One-Shot Video Object Segmentation
- A Benchmark Dataset and Evaluation Methodology for Video Object Segmentation
- Video Object Segmentation without Temporal Information
- Fast Video Object Segmentation by Reference-Guided Mask Propagation
CV-related RC with the lowest percentage of CV-related publications:
With 617 papers between 2015 and 2021, RC 54261 focuses on real-time computing, specifically using distributed computing on the cloud for mobile devices. Papers in this RC range from improvements to real-time learning to mobile augmented reality.4 RC 54261 is a cross-disciplinary subset of research, as it contains scientific publications from different research areas (e.g., distributed computing, real-time computing, computer vision). While computer vision is a contributing area of research to this RC, with 25 percent of papers being CV-related, RC 54261 is not strictly CV research.
RC 54261 Top Five Core Papers:
- Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge
- MCDNN: An Approximation-Based Execution Framework for Deep Stream Processing Under Resource Constraints
- DeepMon: Mobile GPU-based Deep Learning Framework for Continuous Vision Applications
- DeepDecision: A Mobile Deep Learning Framework for Edge Video Analytics
- Chameleon: scalable adaptation of video analytics
CV-related RC in Mathematics
With 217 papers between 2015 and 2021, RC 87188 focuses on noise removal in images, using Possanian image deconvolution and Cauchy noise removal. RC 87188 has 66 percent CV-related publications. As a mathematics-dominant RC, papers in this cluster focus on mathematical approaches to computer vision.
RC 87188 Top Five Core Papers:
- Multiplicative noise removal in imaging: An exp-model and its fixed-point proximitied algorithm
- A new variational approach for restoring images with multiplicative noise
- A convex total generalized variation regularized model for multiplicative noise and blur removal
- Multiplicative noise removal via using nonconvex regularizers based on total variation and wavelet frame
- Cauchy Noise Removal by Nonconvex ADMM with Convergence Guarantees
CV-related RC in Social Science
With 518 papers between 2015 and 2021, RC 5508 focuses on face-processing systems as learned by primates. RC 55088 has 49 percent CV-related publications. As a social science-dominant RC, papers in this cluster explore facial processing and recognition in primates from lab experiments. Because this RC focuses on an important CV topic, facial recognition, this RC gets associated with CV even though many papers are not direct CV implementations.
RC 55088 Top Five Core Papers:
- A Revised Neural Framework for Face Processing
- The Code for Facial Identity in the Primate Brain
- Anatomical Connections of the Functionally Defined “Face Patches” in the Macaque Monkey
- What can we learn about human individual face recognition from experimental studies in monkeys?
- Transformation of Visual Representations Across Ventral Stream Body-selective Patches
Parts three and four of this snapshot mini-series will explore RCs labeled as “NLP” and then Robotics.”
In August 2021, CSET updated the Map of Science, linking more data to the research clusters and implementing a more stable clustering method. With this update, research clusters were assigned new IDs, so the cluster IDs reported in this Snapshot will not match IDs in the current Map of Science user interface. If you are interested in knowing which clusters in the updated Map are most similar to those reported here, or have general questions about our methodology or want to discuss this research, you can email cset@georgetown.edu.
Download Related Data Brief
- Autumn Toney, “Defining Computer Vision, Natural Language Processing, and Robotics Research Clusters ” (Center for Security and Emerging Technology: August 2021).
- Svoboda, E. (2020). Artificial intelligence is improving the detection of lung cancer. Nature, 587(7834), S20–S22. https://doi.org/10.1038/d41586-020-03157-9
- See for example: Perazzi, Federico, et al. “A benchmark dataset and evaluation methodology for video object segmentation.” Proceedings of the IEEE conference on computer vision and pattern recognition. 2016 and Cao, Xiaochun, et al. “Unsupervised pixel-level video foreground object segmentation via shortest path algorithm.” Neurocomputing 172 (2016): 235-243.
- See for example: Lee, Kyungmin, et al. “Outatime: Using speculation to enable low-latency continuous interaction for mobile cloud gaming.” Proceedings of the 13th Annual International Conference on Mobile Systems, Applications, and Services. 2015. And Jain, Puneet, Justin Manweiler, and Romit Roy Choudhury. “Overlay: Practical mobile augmented reality.” Proceedings of the 13th Annual International Conference on Mobile Systems, Applications, and Services. 2015.