![](https://img-blog.csdnimg.cn/20201014180756927.png?x-oss-process=image/resize,m_fixed,h_64,w_64)
笔记
文章平均质量分 88
Ah丶Weii
...
展开
-
[FGD] Focal and Global Knowledge Distillation for detectors (CVPR. 2022)
.原创 2022-07-17 22:37:19 · 1358 阅读 · 1 评论 -
FreeSOLO: Learning to Segment Objects without Annotations* (CVPR. 2022)
freesolo原创 2022-07-17 22:27:33 · 1109 阅读 · 2 评论 -
DCNet (CVPR. 2021)
1. Motivation这篇文章中,作者针对图1出现的2个问题,指出,由于support 和query img 之间的关系没法完全提取,因为之前的方法都是使用GAP的方法,没有考虑局部信息。其次,对于分类估计错误,以及遮挡的问题,Firstly, relations between support fea- tures and query feature are hardly fully explored in previous few-shot detection works, where gl.原创 2021-09-22 15:49:11 · 2065 阅读 · 0 评论 -
[OLN] Learning Open-World Object Proposals without Learning to Classify
1. MotivationOur main insight is that the classifiers in existing object proposers or class agnostic detectors impedes such generalization, because the model tends to overfit to labeled objects and treat the unlabeled objects in the training set as ba.原创 2021-09-06 09:54:34 · 949 阅读 · 0 评论 -
[DeFRCN] Decouple Faster R-CNN for Few-Shot Object Detection(ICCV 2021)
1. Motivation 本文基于Transfer-Learning Based 以及 Faster R-CNN进行改进。 本文针对分类和回归任务中存在的矛盾点进行分析:In this paper, we look closely into the conventional Faster R-CNN and analyze its contradictions from two orthogonal perspectives, namely multi-stage (RPN vs. RCNN.原创 2021-09-02 23:15:11 · 1436 阅读 · 1 评论 -
[FSCE]FSCE: Few-Shot Object Detection via Contrastive Proposal Encoding(CVPR. 2021)
1. Motivation本文是基于fine-tuning based方法In this work, we observe and address the essential weakness of the fine- tuning based approach – constantly mislabeling novel in- stances as confusable categories, and improve the few-shot detection performance to t.原创 2021-08-25 21:48:49 · 2203 阅读 · 0 评论 -
[Retentive R-CNN] Generalized Few-Shot Object Detection without Forgetting(CVPR. 2021)
1. Motivation本文关注于fine-tune后的FSOD模型会在base classes上性能下降的问题。这篇文章构建了Retentive R-CNN,创新点在于Bias-Balance RPN Re-detector,用来在识别novel classes的同时,不降低原有的base classes的精度However, the majority focus merely on the performance of few-shot categories and ignore the ca.原创 2021-08-18 11:51:10 · 835 阅读 · 0 评论 -
Few-Shot Object Detection via Classification Refinement and Distractor Retreatment(CVPR. 2021)
1. MotivationThe current state-of-the-art approach TFA [17] is still far away from satisfaction compared with those general data-abundant detection tasksGiven the fact that TFA is IOU-aware but less semantic discriminative, our key insight is to enh.原创 2021-08-09 21:08:58 · 735 阅读 · 0 评论 -
[TFA] Frustratingly Simple Few-Shot Object Detection(ICML. 2020)
1. Contribution 分类任务上的few-shot研究较多,相比之前FSOD收到较少的关注。Detecting rare objects from a few examples is an emerging problem.However, much of this work has focused on basic image classification tasks. In contrast, few-shot object detection has received .原创 2021-08-09 13:05:57 · 1539 阅读 · 2 评论 -
[SRR-FSD] Semantic Relation Reasoning for Shot-Stable Few-Shot Object Detection(CVPR. 2021)
[外链图片转存中…(img-7uQh5Zw9-1628130409148)]1. Motivationfew shot 本身存在的意义:In other words, we are unable to alleviate the situation of scarce cases by simply spend- ing more money on annotation even big data is accessible.Therefore, the study of few-shot原创 2021-08-07 22:00:24 · 700 阅读 · 0 评论 -
Probablitic two-stage detection
1. MotivationWhile the second stage has a probabilistic interpretation, the combination of the two stages does not.A probabilistic two-stage detector is faster and more accu- rate than both its one- and two-stage precursors.2. ContributionWe build .原创 2021-08-02 21:06:48 · 230 阅读 · 0 评论 -
[OVD]Open-Vocabulary Object Detection Using Captions(CVPR. 2021 oral)
1. MotivationDespite the remarkable accuracy of deep neural networks in object detection, they are costly to train and scale due to supervision requirements.Weakly supervised and zero-shot learning techniques have been explored to scale object detec.原创 2021-07-21 22:20:08 · 1175 阅读 · 0 评论 -
[Det-AdvProp] Robust and Accuracy Object Detection via Adversarial Learning(CVPR. 2021)
1. Motivation数据增强在分类网络中应用广泛,但是在目标检测中under-explored.Data augmentation has become a de facto component for training high-performance deep image classifiers, but its potential is under-explored for object detection.In this paper, we aim to enhance this l.原创 2021-07-21 17:21:30 · 707 阅读 · 0 评论 -
Cycle GAN(ICCV. 2017)
1. MotivationFor many tasks, paired training data will not be available.We present an approach for learning to translate an image from a source domain X to a target domain Y in the absence of paired examples.对于Cycle consistent的引入很重要,否则会出现模式崩溃的问题,所有输入.原创 2021-06-30 20:31:15 · 129 阅读 · 0 评论 -
Pix2Pix GAN(CVPR. 2017)
1. MotivationImage-to-Image translation的定义We define automatic image-to-image translation as the task of translating one possible representation of a scene into another.Our goal in this paper is to develop a common framework for all these problems.需要.原创 2021-06-28 22:41:34 · 700 阅读 · 0 评论 -
GAN[NIPS. 2014]
KL divergence,如果KL的值越大,代表2个分布之间的差异越大,KL的值跃萧,代表2个分布之间的差异最小。KL divergence:DKL(P∣∣Q)=∑i=1NP(xi)logP(xi)Q(xi)D_{KL}(P||Q) = \sum^N_{i=1} P(x_i)log\frac{P(x_i)}{Q(x_i)}DKL(P∣∣Q)=i=1∑NP(xi)logQ(xi)P(xi)JS divergence:JSD(P∣∣Q)=12D(P∣∣M)+12D(Q∣∣M)M=1..原创 2021-06-23 17:05:12 · 104 阅读 · 0 评论 -
[detectron2 ] Mask R-CNN代码笔记
RPNRPN中的`match_label`表示的是[0,-1,1]# forward pairwise_iou functionmatch_quality_matrix = retry_if_cuda_oom(pairwise_iou)(gt_boxes_i, anchors) # [M x N] for each img# forward match functionmatched_idxs, gt_labels_i = retry_if_cuda_oom(self.anchor_matche原创 2021-06-21 11:41:18 · 2065 阅读 · 0 评论 -
[PolarMask++](TPAMI. 2021)
1. ContributionThe main contributions of this work are three-fold:We introduce a new perspective to design a single-shot instance segmentation framework, PolarMask, which predicts instance masks and rotated objects in the polar coordinate in an effecti.原创 2021-06-15 11:37:33 · 518 阅读 · 0 评论 -
[SOLQ]SOLQ: Segmenting Objects by Learning Queries
1. MotivationAnchor-based segmentation: the segmentation branch heavily relies on the detection branch, making it hard to achieve better joint learning of multiple tasks.Anchor-free segmentation: the weights of dynamic convolutions or the mask coeffic.原创 2021-06-15 10:54:36 · 794 阅读 · 0 评论 -
[CenterMask]CenterMask: Real-Time Anchor-Free Instance Segmentation(CVPR. 2020)
1. ContributionWe design a scale-adaptive RoI assignment function that considers the input scale and is a more suitable one-stage object detector.We also propose a more effective backbone network VoVNetV2 based on VoVNet, which shows better performa.原创 2021-06-09 12:15:05 · 229 阅读 · 0 评论 -
[MLAD] Modeling Multi-Lalbel Action Dependencies for Temporal Action Localization(CVPR. 2021 oral)
1. Motivation以往的大部分方法没有显示建模不同action label的关系。Although these works achieve strong multi-label action localization performance, they do not explicitly model the relationships between the different action labels, which can be extremely useful for determ..原创 2021-06-08 10:46:53 · 595 阅读 · 0 评论 -
DyCo3D (CVPR. 2021)
1. MotivationPrevious top-performing approaches for point cloud instance segmentation involve a bottom-up strategy, which often includes inefficient operations or complex pipelines, such as grouping over-segmented components, introducing additional step.原创 2021-05-28 10:54:51 · 887 阅读 · 0 评论 -
[PointNet++] (NPIS 2017)
文章目录1. Motivation2. Contribution3. Method3.1 Hierarchical Point Set Feature Learning3.1.2 Sampling layer3.1.3 Grouping layer3.1.4 PointNet layer3.2 Robust Feature Learning under Non-Uniform Sampling Density3.2.1 Multi-scale grouping (MSG)3.2.2 Multi-reso..原创 2021-05-25 21:38:09 · 247 阅读 · 0 评论 -
[PointNet] PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation(CVPR. 2017)
文章目录1. Motivation2. Introduction and Related Work2.1 Introduction2.2 Related Work3. Contribution4. Method4.1. Properties of Point Sets in RnR^nRn4.2 PointNet Architecture5. Experiment5.1 applications5.1.1 3D Object Classification5.1.2 3D Object Part Segme.原创 2021-05-25 18:21:43 · 234 阅读 · 0 评论 -
[Vscode] 上传Github笔记
1. 初始化Git首先在github上创建一个repository,修改默认的default branch 为master 后续用main会有问题。然后使用终端git(windows10)或者linux 终端。执行以下操作:# 1. cd your uploaded dirgit init # 2. add ignore-filetouch .gitignore# 3. addgit add -A# 4. upload local repositorygit commit -m原创 2021-04-28 17:19:44 · 549 阅读 · 0 评论 -
[python] zip 和 *使用的进一步理解
zip(*)例子,拆解rs,如果使用zip(rs)则是错误的rs = [[1,2,3],[1,2,3],[1,2,3],[1,2,3]]ra = list(zip(*rs)) # correct# ra[(1, 1, 1, 1), (2, 2, 2, 2), (3, 3, 3, 3)]rb = list(zip(rs)) #wrong#rb[([1, 2, 3],), ([1, 2, 3],), ([1, 2, 3],), ([1, 2, 3],)]*print(*rs) # 把原创 2021-04-26 13:43:47 · 69 阅读 · 0 评论 -
[MS R-CNN] Mask Scoring R-CNN(CVPR. 2019 oral)
1. MotivationMask Scoring R-CNN 是在Mask R-CNN的基础之上,由于分类的得分scores没法很好的反映mask的quality,例如有些很高的scores 但是mask的quality却比较差,并且随着cls scores的增大,mask quality并不是呈现一种线性的关系。因此将Smask = Scls * Smask_iou,decompose成2个部分,第一个cls的得分可以直接用faster rcnn的检测分类分支,而maskiou作为本文的一个.原创 2021-04-23 13:09:59 · 249 阅读 · 1 评论 -
[RepVGG] RepVGG:Making VGG-style ConvNets Great Again(CVPR.2021)
paper:https://arxiv.org/abs/2101.0369code:https://arxiv.org/abs/2101.03691. Motivation 如今更复杂的卷积网络可以取得更大的精度,但是相对于简单的卷积网络来说,也有2种缺点,首先是多分支结构中复杂的设定,以及复杂卷积网络的计算资源的开销。 同时,其他simple ConvNets的性能并不能比得过complicated ConvNets。2. Contribution本文的贡献如下:本文提出了Rep..原创 2021-03-04 21:43:41 · 429 阅读 · 0 评论 -
[PSS]Object Detection Made Simpler by Eliminating Heuristic NMS
文章目录1. Motivation2. Contribution3. Our Method3.1 Overall Training Objective3.1.1 PSS LOSS3.1.2 Ranking Loss3.2 One-to-many Label Assignment3.3 One-to-one Label Assignment3.4 Conflict in the Tow Classification Loss Terms3.5 Stop Gradient4. Experiments4.1 A.原创 2021-02-01 17:00:31 · 313 阅读 · 0 评论 -
[Relation Network]Realtion Networks for Object Detection (CVPR. 2018)
[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-nbwrgm7Y-1611391066244)(https://raw.githubusercontent.com/Wei-i/My_Image_Hosting/main/img/image-20210121215141004.png)]1. Motivationintuition: 认为建模物体之间的关系会帮助目标检测。Although it is well believed for years that model原创 2021-01-23 16:39:36 · 251 阅读 · 0 评论 -
[EmbedMask]EmbedMask: Embedding Coupling for One-stage Instance Segmentation
文章目录1. Motivation2. Contribution3. EmbedMask3.1 Overview3.2 Embedding Definition3.3 Learnable Margin3.4 Smooth Loss3.5 Training3.6 Inference1. Motivationtwo-stage 通过ROIPool方法,会造成图像信息的丢失(低分辨率以及ROIPool/Align)的对齐操作。同时它的参数量较多比较复杂。现有(2019)的one-stage实例分割.原创 2021-01-22 17:11:21 · 322 阅读 · 0 评论 -
[Sparse R-CNN]Sparse R-CNN: End-to-End Object Detection with Learnable Proposals (CVPR. 2021)
Sparse R-CNN: End-to-End Object Detection with Learnable Proposalspaper:https://arxiv.org/pdf/2011.12450.pdfcode:https://github.com/PeizeSun/SparseR-CNNAbstract作者提出了Sparse R-CNN,一种对于图像目标检测的sparse方法,一种固定的可学习目标候选框,总数量为N,用于目标识别,来进行分类和定位。通过将H x W x K个手工设定的原创 2020-12-23 10:30:43 · 1336 阅读 · 2 评论 -
[CenterNet]Objects as Points笔记
Objects as Points paper:https://arxiv.org/pdf/1904.07850.pdfcode:https://github.com/xingyizhou/CenterNet/环境配置参考:官方安装文档,很棒的教学博客1. Motivation大部分的目标检测网络都需要详尽的可能的目标定位以及分类的anchor,这是非常浪费,不高效并且要引入预处理操作,如NMS。在这篇论文中,作者提出了一个不同的方法,成为CenterNet,将物体建模为a single po.原创 2020-12-22 10:04:11 · 129 阅读 · 0 评论 -
[OneNet]OneNet: Towards End-to-End One-Stage Object Detection笔记
paper:https://arxiv.org/pdf/2012.05780.pdfcode:https://github.com/PeizeSun/OneNet1. 摘要:文中提出了一个观点,认为以往的one-stage无法去掉NMS实现端到端的主要原因在于正样本的选取。以往的Label assignment任务没有考虑分类的cost,只考虑了位置的cost,从而会得到很多冗余的boxes,造成了后处理中必须使用NMS。如下图所示,RetinaNet选取正样本只考虑了box的IoU ,而FCOS原创 2020-12-18 21:17:48 · 449 阅读 · 0 评论 -
[DETR] End-to-End Object Detection with Transformers (ECCV. 2020 oral)代码笔记
End-to-End Object Detection with Transformers文章目录End-to-End Object Detection with Transformers网络结构detr/models/detr.pydetr/models/backbone.py论文:https://arxiv.org/pdf/2005.12872.pdf代码:https://github.com/facebookresearch/detr网络结构detr/models/detr.py代码原创 2020-12-08 19:31:39 · 1130 阅读 · 0 评论 -
[Condinst]Conditional Convolutions for Instance Segmentation(ECCV. 2020 oral)
文章目录网络结构mask headLOSS1. AdelaiDet/adet/modeling/condinst/condinst.py2. AdelaiDet/adet/modeling/condinst/mask_branch.py3. AdelaiDet/adet/modeling/condinst/dynamic_mask_head.py4. AdelaiDet/adet/modeling/fcos/fcos_outputs.py中Condinst的top_feat结构网络结构mask hea原创 2020-12-04 23:32:59 · 1515 阅读 · 3 评论 -
[BlendMask]BlendMask: Top-Down Meets Bottom-Up for Instance Segmentation代码笔记
文章目录BlendMask 网路结构:1. AdelaiDet/adet/modeling/blendmask/blendmask.py2. AdelaiDet/adet/modeling/blendmask/blender.py3. AdelaiDet/adet/modeling/blendmask/basis_module.py总的执行顺序为 backbone fpn and resnet --> fcos --> blendmask.py --> basis_module.py原创 2020-11-22 22:58:56 · 1544 阅读 · 3 评论 -
[FCOS]FCOS: Fully Convolutional One-Stage Object Detection(ICCV. 2019)代码笔记
文章目录1. AdelaiDet/adet/modeling/fcos/fcos.py2. AdelaiDet/adet/modeling/fcos/fcos_outputs.py3. AdelaiDet/adet/layers/iou_loss.py1. AdelaiDet/adet/modeling/fcos/fcos.pyimport mathfrom typing import List, Dictimport torchfrom torch import nnfrom torch.n原创 2020-11-21 20:34:34 · 681 阅读 · 2 评论 -
[SOLO ]SOLO: Segmenting Objects by Locations代码解读笔记(ECCV. 2020)
文章目录SOLO head网络结构损失函数1. SOLO/mmdect/models/detectors/single_stage_ins.py2. SOLO/mmdet/models/anchor_heads/solo_head.py3. SOLO/mmdetect/core/post_processing/matrix_nms.py4. SOLO/configs/solo/solo_r50_fpn_8gpu_1x.py5. SOLO/mmdet/models/anchor_heads/_ _init_原创 2020-11-20 19:34:37 · 1676 阅读 · 3 评论 -
[实例分割]IOU,NMS笔记
1.IOU参考:知乎自己没法直接理解多维度多个box之间的iou,索性一步步写下来。pytoch 源代码=# IOU计算 # 假设box1维度为[N,4] box2维度为[M,4] def iou(self, box1, box2): N = box1.size(0) M = box2.size(0) lt = torch.max( # 左上角的点 注意:是inter的左上角的点,因此才要算MAX b原创 2020-11-10 20:30:01 · 1308 阅读 · 0 评论