Wacv 2021 paper large disk – image segmentation


Compile | CV
Report | opencv Chinese website (wechat ID: iopencv)

This article reviews the papers related to wacv2021 image segmentation, including matting, instance, panorama, semantic segmentation, natural disaster assessment and other related applications. It is noteworthy that there is a text matting that has been rarely or never studied in previous work.

A total of 11 articles. If there is any omission, please add.

Download all wacv 2021 papers containing these papers:

“Wacv 2021 opening, more emphasis on technology application, with paper Download”

Instance segmentation

[1].Weakly Supervised Instance Segmentation by Deep Community Learning

In this paper, the author introduces a deep community learning framework for weakly supervised instance segmentation. The framework is based on an end-to-end trainable deep neural network, which has active interaction among multiple tasks of target detection, instance mask generation and target segmentation. Two experienced target location technologies are added: class agnostic bounding box region and segmentation proposal generation, which are carried out without complete supervision.

Without post-processing, the proposed algorithm achieves significantly higher performance than the existing weak supervision methods on the standard benchmark data set.

Author|   Jaedong Hwang, Seohyun Kim, Jeany Son, Bohyung Han

Unit|   Seoul University; ETRI

Thesis|   https://arxiv.org/abs/2001.11207

Home page|   https://cv.snu.ac.kr/research…

Wacv 2021 paper large disk - image segmentation

[2].MSNet: A Multilevel Instance Segmentation Network for Natural Disaster Damage Assessment in Aerial Videos

Msnet: multi-level instance segmentation network for aerial video natural disaster assessment

This paper studies the effective evaluation of building losses after natural disasters such as hurricanes, floods or fires through aerial video analysis.

Made two main contributions:

The first contribution is a new data set, including aerial videos generated by social media users with instance level building damage mask annotation. It provides the first benchmark for the quantitative evaluation of building damage model using aerial video.

The second contribution is a new model: msnet, which includes a new regional proposal network design and an unsupervised fractional refinement network for confidence score calibration of boundary box and mask branches.

Experiments show that the new model and new data set achieve the most advanced results compared with the previous methods.

The authors said they would publish data, models and code.

Author | Xiaoyu Zhu, Junwei Liang, Alexander Hauptmann

Unit: Carnegie Mellon University

Thesis| https://w.url.cn/s/AC2eIwl

Code| https://github.com/zgzxy001/M…

Wacv 2021 paper large disk - image segmentation

Semantic segmentation

[3].Unsupervised Domain Adaptation in Semantic Segmentation via Orthogonal and Clustered Embeddings

Author|   Marco Toldo,   Umberto Michieli,   Pietro Zanuttigh

Unit|   University of Padova

Thesis|   https://arxiv.org/abs/2011.12616

Home page|   https://lttm.dei.unipd.it/pap…

Code|   https://github.com/LTTM/UDAcl…

Wacv 2021 paper large disk - image segmentation

[4].Improving Point Cloud Semantic Segmentation by Learning 3D Object Detection

Detection aware 3D semantic segmentation (DASS) network is proposed to solve the limitations of the current architecture.

Dass can improve the 3D semantic segmentation result of geometric similarity class to 37.8% IOU of image FOV while maintaining the high-precision aerial view (Bev) detection result.

Author|   Ozan Unal,   Luc Van Gool,   Dengxin Dai

Unit|   Zurich Federal Institute of technology; University of Leuven

Thesis|   https://arxiv.org/abs/2009.10569

Wacv 2021 paper large disk - image segmentation

[5].Multi Projection Fusion for Real-Time Semantic Segmentation of 3D LiDAR Point Clouds

In this work, the author proposes a new multi projection fusion framework, which uses spherical and aerial projection and soft voting mechanism to fuse the results to realize the semantic segmentation of point cloud. The Miou of the proposed framework on semantickitti dataset reaches 55.5, which is higher than the most advanced methods based on single projection, rangenet + + and polarnet, 1.6 times faster than the former and 3.1 times faster than the latter. At the same time, it has higher throughput.

For future work, the first mock exam is to combine two projections into a single multi view unified model, and to study more than two projections in the framework.

Author|   Yara Ali Alnaggar,   Mohamed Afifi,   Karim Amer,   Mohamed Elhelw

Unit: Nile University;

Thesis|   https://arxiv.org/abs/2011.01974

Wacv 2021 paper large disk - image segmentation

[6].Shape From Semantic Segmentation via the Geometric Renyi Divergence

By | Tatsuro Koizumi, William A. P. Smith

Unit|   York University



Wacv 2021 paper large disk - image segmentation

Panoramic segmentation

[7].Boosting Monocular Depth With Panoptic Segmentation Maps

Author|   Faraz Saeedan, Stefan Roth

Unit|   Darmstadt University of Technology

Thesis|   https://openaccess.thecvf.com…

Wacv 2021 paper large disk - image segmentation

Video object segmentation

[8].Reducing the Annotation Effort for Video Object Segmentation Datasets

In order to further improve the performance of video target segmentation, larger, more diverse and more challenging data sets are needed. However, dense labeling of each frame with pixel mask can not be extended to large data sets.

Therefore, the author uses the deep convolution network to automatically create pixel level pseudo tags from cheaper bounding box annotation, and studies to what extent this pseudo tag can carry the most advanced Vos method of training. It is gratifying that only adding a manually labeled mask to a single video frame of each object is enough to generate a pseudo label to train the Vos method, and achieve almost the same performance level as that of fully segmented video training.

Based on this, a pixel pseudo label is created for the training set of Tao data set, and a subset of the verification set is manually labeled. The new tao-vos benchmark is obtained and published in https://www.vision.rwth-aache… (recently published)

By Paul Voigtlaender, Lishu Luo, Chun Yuan, Yong Jiang, Bastian leibe

Unit: Aachen University of technology; Tsinghua University

Thesis| https://arxiv.org/abs/2011.01142

Code| https://www.vision.rwth-aache…

Remarks | wacv 2021

Wacv 2021 paper large disk - image segmentation

Target partial segmentation

[9].Part Segmentation of Unseen Objects using Keypoint Guidance

The author develops an end-to-end learning method, which uses the key point position to guide the transfer learning process, and migrates the pixel level target partial segmentation from the fully labeled target set to another weakly labeled target set. For partial segmentation, the author proves that the nonparametric template matching method is more effective than pixel classification, especially for small or infrequent parts.

In order to verify the generality of the proposed method, the author introduces two new data sets, including 200 quadrupeds, with key points and partial segmentation annotation. It is proved that the proposed method uses limited partial segmentation tags in the training process, and can be superior to the existing models in the task of new object partial segmentation.

Author | shujon Naha, Qingyang Xiao, prianka Banik, MD. alimoor Reza, David J. Crandall

Unit|   Indiana University School of Arts and Sciences

Thesis|   https://openaccess.thecvf.com…

Data set|   http://vision.sice.indiana.ed…

Wacv 2021 paper large disk - image segmentation

image matting

Towards Enhancing Fine-Grained Details for Image Matting

A new viewpoint on image matting is proposed, which is clearly divided into two parts: one is the semantic part of extracting high-level semantic clues, and the other is the texture compensation part of providing fine details and low-level texture clues.

Based on this, a new depth image matting method is proposed, which clearly defines two paths: encoder decoder semantic path and no down sampling texture compensation path. A new loss term is further proposed to help the network alleviate the inaccurate trimap problem and better detect those “pure” background parts.

The proposed method achieves new state-of-the-art performance on the challenging Adobe composition 1K test data set.

Author|   Chang Liu, Henghui Ding, Xudong Jiang

Unit|   Nanyang University of Technology

Thesis| https://openaccess.thecvf.com…

Wacv 2021 paper large disk - image segmentation

Text Matting

ATM: Attentional Text Matting

In this study, the author tries to solve the text matting problem of extracting characters (usually wordarts) from the image background. Different from the traditional image matting problem, text matting is much more difficult, because its prospect has three characteristics: small, multi-objective, complex structure and boundary.

The scheme is to propose a two-order attention text matting pipeline, which the author calls the first text matting method. A text matting image synthesis engine is constructed, and a large-scale high-quality text matting data set with diversity is synthesized.

A large number of experiments on synthetic and real image data sets show that the proposed method is superior to the most advanced image matting method in text matting task.

By Peng Kang,   Jianping Zhang,   Chen Ma,   Guiling Sun

Setting: Northwestern University; McGill University, Canada; Nankai University


Code|   https://github.com/TextMattin…

Wacv 2021 paper large disk - image segmentation

Wacv 2021 paper large disk - image segmentation

[recommended reading]

Wacv 2021 paper large market point – target detection

Wacv 2021 paper – face technology

– END –