Discussion on several research hotspots of cvpr2022

CVPR2022 Just finished , As the most influential visual event , This year, another batch of excellent work was displayed . I believe that all of you who pay attention to the latest research progress in vision , Have sharpened their fists , Prepare to CVPR2023 Submitted . Based on this year's work , Which fields are CVPR Focus of attention ？ What areas of work , Higher acceptance ,oral The proportion is larger ？ be based on CVPR The latest official statistics , I will talk to you CVPR Some research hotspots of , Hope to cast a round on those plans CVPR Of students to provide a little reference information .

---

1. Ten hot research fields

First , We are based on oral Statistical information of the paper , Sort according to the proportion of papers received and the fields mentioned , Get ten hot areas , Include ： Multi angle 3D vision , Image and video synthesis , Identification, detection, classification and retrieval , Deep network structure design , Vision and language processing intersect , Visual analysis of low quality data , Shape analysis , The migration study , Video analysis and understanding , Attitude estimation .

chart 1. Ten hot research areas （Oral）

When we count all received papers , The statistics will change a little in order , Include ： Identification, detection, classification and retrieval , Image and video synthesis , Multi angle 3D vision , Visual analysis of low quality data , Vision and language processing intersect , Shape analysis , The migration study , Deep network structure design , Self supervised and unsupervised learning , Video analysis and understanding .

chart 2. Ten hot research areas （All）

You can see , The two sorts correspond to the hot research issues , High repeatability . Combine two tables , Focus on the application level to summarize the hot spots , I have selected five hot research areas , For students who plan to contribute ：

• Multi angle 3D vision
• Image and video synthesis
• Identification, detection, classification and retrieval
• Vision and language processing intersect
• Visual analysis of low quality data

---

2. Best Paper

CVPR2022 Of Best paper list Contains four articles , Respectively ：

Best Paper Award: Learning to Solve Hard Minimal Problems

Best Paper Honorable Mention: Dual-Shutter Optical Vibration Sensing

Best Student Paper Award: EPro-PnP: Generalized End-to-End Probabilistic Perspective-n-Points for Monocular Object Pose Estimation

Best Student Paper Honorable Mention: Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields

The best paper is 《Learning to Solve Hard Minimal Problems》. Took a rough look , Is not very good , I have done some theoretical work in the field of optimization problems , Some tools of geometric optimization are introduced .《Dual-Shutter Optical Vibration Sensing》 It's about three-dimensional laser scanning technology .《EPro-PnP: Generalized End-to-End Probabilistic Perspective...》 Based on multipoint perspective theory , This paper presents a method to estimate the 3D pose of an object from an image .《Ref-NeRF》 Basically is NeRF Research on variants of Algorithm . From the emphasis of the best paper, we can know ,CVPR Prefer 3D vision related research . in addition , Those who had a high voice before the meeting Kaiming The teacher's 《Masked Autoencoders Are Scalable Vision Learners》 It is also worth studying deeply . be based on MAE This paper proposes a method based on patch Predicted codec structure , It has excellent prediction and reconstruction performance for data image content understanding . The paper was listed as the best candidate for the paper .

---

3. Personal concerns

Because I have been doing color migration recently , Lighting optimization and other work , So I pay more attention to low-level vision field . This year, CVPR I have enrolled in this field 19 piece oral as well as 91 piece poster, The number of articles received cannot be counted as small . I will be corresponding to 19 piece oral The article is copied here , After convenient study .

[1] Robust Equivariant Imaging: A Fully Unsupervised Framework for Learning To Image From Noisy and Partial Measurements. （ Denoise + Super resolution is used in image enhancement technology ）

[2] Bijective Mapping Network for Shadow Removal. （ Eliminate shadows ）

[3] Event-Aided Direct Sparse Odometry. ( Sparse point cloud enhancement )

[4] MAXIM: Multi-Axis MLP for Image Processing.（ General image quality enhancement algorithm ）

[5] Details or Artifacts: A Locally Discriminative Learning Approach to Realistic Image Super-Resolution.（ Super resolution ）

[6] Dual Adversarial Adaptation for Cross-Device Real-World Image Super-Resolution. （ Super resolution ）

[7] ELIC: Efficient Learned Image Compression With Unevenly Grouped Space-Channel Contextual Adaptive Coding. 

[8] Discrete Cosine Transform Network for Guided Depth Map Super-Resolution. ( Super resolution )

[9] Deep Rectangling for Image Stitching: A Learning Baseline.（ Image mosaic ）

[10] CamLiFlow: Bidirectional Camera-LiDAR Fusion for Joint Optical Flow and Scene Flow Estimation. ( Optical flow optimization )

[11] Toward Fast, Flexible, and Robust Low-Light Image Enhancement. ( Low light enhancement )

[12] Faithful Extreme Rescaling via Generative Prior Reciprocated Invertible Represe-ntations.

[13] Learning Trajectory-Aware Transformer for Video Super-Resolution. ( Super resolution )

[14] SphereSR: 360deg Image Super-Resolution With Arbitrary Projection via Continuous Spherical Image Representation.（ Super resolution ）

[15] Parametric Scattering Networks. ( Optimized learning structure )

[16] Target-Aware Dual Adversarial Learning and a Multi-Scenario Multi-Modality Benchmark To Fuse Infrared and Visible for Object Detection. ( Object detection in low light environment )

[17] Learning to Deblur Using Light Field Generated and Real Defocus Images. ( To blur )

[18] Burst Image Restoration and Enhancement. ( Image reconstruction )

[19 ]Restormer: Efficient Transformer for High-Resolution Image Restoration. ( To blur )

stay low-level vision field , Super resolution still accounts for a large proportion . Some of the work involves de obfuscation , Quality enhancement , Detail reconstruction, etc , In essence, it is closely related to super-resolution technology . It seems , What to do in the future low-level vision, The probability should use the super-resolution algorithm . It can be seen from some papers that , Three dimensional vision has been integrated into low-level vision field . For depth map , Panoramic photos and other data with three-dimensional attributes , Detail reconstruction , Motion compensation, etc , It is also a very good research direction .