Paper List - WangShaoRu/paper-reading GitHub Wiki

Paper List

2020.4.30

  • NMS by Representative Region : Towards Crowded Pedestrian Detection by Proposal Pairing
    Date: 2020.3.28
    Topic: Crowded Pedestrian Detection
    Rating: $\star\star\star$
    Intro: This paper proposes a novel Representative Region NMS (R2NMS) approach leveraging the less occluded visible parts, effectively removing the redundant boxes without bringing in many false positives. Authors&teams: Xin Huang(Waseda University)
    Code:
    Contributions&Methods:

    • a novel NMS method – R2NMS, to overcome the weakness of original NMS.
    • a Paired-Box Model (PBM) which simultaneously predicts both the full and visible boxes of a single pedestrian, and performs convenient feature integration of the two boxes.

    Introducer: LiZekun

2020.4.30

  • Stitcher: Feedback-driven Data Provider for Object Detection
    Date: 2020.4.26
    Topic: Object Detection
    Rating: $\star\star\star$
    Intro: Stitcher, a feedback-driven data provider, which aims to train object detectors in a balanced way
    Authors&teams: Yukang Chen(NLPR)
    Code:
    Contributions&Methods:

    • We propose Stitcher, a feedback-driven data provider, that enhances the performance of object detection, by utilizing training loss in a feedback manner.

    Introducer: LiZekun

2020.3.27

  • SOLOv2: Dynamic, Faster and Stronger
    Date: 2019.3.23
    Topic: Object Detection
    Rating: $\star\star\star$
    Intro: Improve SOLO from mask learning and mask NMS
    Authors&teams: Xinlong Wang, Tao Kong(The University of Adelaide, ByteDance AI Lab)
    Code: git.io/AdelaiDet
    Contributions&Methods:

    • Mask branch is decoupled into a mask kernel branch and a mask feature branch. And then the mask is predicted by convolving the mask kernel on the mask feature.
    • Matrix NMS: Constuct a decay factor for each prediction to approximate the suppression in NMS and achieve parallel computation.

    Introducer: Yangli

  • Detection in Crowded Scenes: One Proposal, Multiple Predictions
    Date: 2019.3.20
    Topic: Object Detection
    Rating: $\star\star\star$
    Intro: Multiple predictions from one proposal for detection in crowded scenes
    Authors&teams: Xuangeng Chu, Jian Sun(Peking University, MEGVII Technology)
    Code:
    Contributions&Methods:

    • For each proposal, predict a set of instances instead of a single instance.
    • EMD loss is proposed to minimize the gap between the predictions and the set of matched gt instances.
    • Set NMS: Skip the suppresstion between the predictions from the same proposal.

    Introducer: Yangli

2020.3.20

  • PointINS: Point-based Instance Segmentation
    Date: 2019.3.13
    Topic: Object Detection
    Rating: $\star\star\star$
    Intro: Instance segmentation using instance-aware convolution
    Authors&teams: Lu Qi, Jian Sun, Jiaya Jia (The Chinese University of Hong Kong, MEGVII Technology, SmartMore)
    Code:
    Contributions&Methods:

    • The instance-agnostic feature is generated from single-point feature by channel up-scaling and depth-to-space. The instance-aware weight is obtained by transforming the predicted bbox information. Convolve the instance-agnostic feature with the instance-aware weight for mask prediction.

    Introducer: Yangli

2020.3.13

  • CPM R-CNN: Calibrating Point-guided Misalignment in Object Detection
    Date: 2019.3.7
    Topic: Object Detection
    Rating: $\star\star$
    Intro: Several modules to improve Grid RCNN
    Authors&teams: Bin Zhu (Beijing University of Posts and Telecommunications)
    Code:
    Contributions&Methods:

    • Cascade mapping module: Apply cascade stages with coarse-to-fine mapping ratios to refine proposals.
    • IoU scoring module & Resampling scoring module: Fuse cls scores with predicted IoUs and outputs of a classifier trained with refined RoIs & negative RoIs.

    Introducer: Yangli

2020.01.03

  • Scale Match for Tiny Person Detection

    Date: 2019.12.23
    Topic: Object Detection
    Rating: $\star\star$
    Intro: A new benchmark of tiny person
    Authors&teams: Xuehui Yu (UCAS)
    Code:https://github.com/ucas-vg/TinyBenchmark

    Contributions&Methods:

    • Tiny person benchmark: The absolute size 18.0 which is 99.5 in coco of objects in dataset.
    • Scale Match: A efficient scale transformation approach for tiny person detection by keeping the scale consistency between the TinyPerson and the extra dataset.

2019.12.31

  • AugFPN: Improving Multi-scale Feature Learning for Object Detection
    Date: 2019.12.11
    Topic: Object Detection
    Rating: $\star\star$
    Intro: A new pyramid structure based on FPN
    Authors&teams: Chaoxu Guo (NLPR)
    Code:
    Contributions&Methods:

    • Consistent Supervision: There is an additional predict layer between the bottom-top and top-bottom layer.
    • Residual Feature Augmentation: A parallel structure is proposed to combine the features from different scales based on channel or spatial.

    Introducer: LiZekun

  • Dually Supervised Feature Pyramid for Object Detection and Segmentation
    Date: 2019.12.08
    Topic: Object Detection
    Rating: $\star\star$
    Intro: A new method for FPN and classification and location
    Authors&teams: Fan Yang、Haibin Ling (Temple University, Philadelphia, USA)
    Code:
    Contributions&Methods:

    • An additional predict: There is an additional predict layer between the bottom-top and top -bottom layer.
    • Decoupled head : The decoupled head separates the classification and regression tasks in hidden feature space, which is achieved by taking apart the shared two hidden layers with.

    Introducer: LiZekun

2019.12.27

  • Side-Aware Boundary Localization for More Precise Object Detection
    Date: 2019.12.9
    Topic: Object Detection
    Rating: $\star\star\star$
    Intro: A new alternative approach for bbox localization
    Authors&teams: Jiaqi Wang (cuhk, The Chinese University of Hong Kong)
    Code: https: //github.com/open-mmlab/mmdetection
    Contributions&Methods:

    • Side-aware Feature Extraction: self-attention mechanism and normalization is applied to each ROI feature to get two vector with shape: 1*k and k*1.
    • Boundary Localization With Bucketing: vectors obtained by Side-aware Feature Extraction are then used to classify (whether each pixel is boundary) and regress (where is the boundary with reference to this pixel).

    Introducer: Wangsr

2019.12.19

  • Multiple Anchor Learning for Visual Object Detection
    Date: 2019.12.4
    Topic: Object Detection
    Rating: $\star\star\star$
    Intro: A new stategy for anchor-object matching and optimization
    Authors&teams: Wei Ke (CMU)
    Code:
    Contributions:

    • Anchor selection: Construct an anchor bag for each gt object according to IoUs. Select top-scored anchors from each bag for learning, and the selection proportion is decreased with iterations.
    • Anchor depression: To avoid choosing sub-optimal anchors, perturb the features of selected anchors to decrease their confidences.

    Introducer: Yangli

  • Mixture-Model-based Bounding Box Density Estimation for Object Detection
    Date: 2019.11.28
    Topic: Object Detection
    Rating: $\star\star\star$
    Intro: A mixture model to capture the distribution of bboxes
    Authors&teams: Jaeyoung Yoo (Seoul National University)
    Code:
    Contributions:

    • Propose a mixture model to capture the distribution of bbox coordinates.
    • Sample RoIs from the mixture model to learning class probability (including background).

    Introducer: Yangli

  • Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression
    Date: 2019.11.19
    Topic: Object Detection
    Rating: $\star\star$
    Intro: Improvements to IoU and GIoU losses
    Authors&teams: Zhaohui Zheng (Tianjin University)
    Code: https://github.com/Zzh-tju/DIoU
    Contributions:

    • To achieve faster convergence, Distance-IoU loss adds normalized distance between two bboxes to IoU loss.
    • Complete IoU loss further considers the consistency of aspect ratios for two bboxes.
    • Deployed in NMS, DIoU loss is more robust than original NMS.

    Introducer: Yangli

2019.11.28