K-Net:Towards Unified Image Segmentation Posted on 2022-05-20 | Visitors: Words count in article: 2,508 | Reading time ≈ 9 K-Net 统一了语义、实例和全景分割框架,它通过一组可学习的 kernels 来分割实例和语义类别,其中每个 kernel 负责为潜在实例或 stuff 类别生成 mask。为了解决区分不同实例的困难,论文提出一种 kernel update 策略,改策略使每个 kernel 能够动态并以输入图像中意义组为条件。 Read more »
SOTR:Segmenting Objects with Transformers [ICCV 2021] Posted on 2022-05-19 | Visitors: Words count in article: 2,433 | Reading time ≈ 9 Segmenting Objects with TRansformers (SOTR) 简化了分割的pipeline,它具有2个并行的子任务:(1)通过 transformer 预测每个实例类别,(2)使用多层级上采样模块动态生成 segmentation mask。SOTR 可以分别通过特征金字塔(FPN)和 twin transformer 有效地提取较低级别的特征表示(lower-level feature representations)并不惑远程上下文依赖关系(long-range context dependencies)。同时,与原始的 tranformer 相比,多提出的 twin transformer 在时间和资源上都是有效的,因为只涉及行和列注意力(a row and a column attention )来编码像素。 Read more »
Efficient DETR Posted on 2021-12-27 | In Object Detection | Visitors: Words count in article: 3,176 | Reading time ≈ 12 Efficient DETR 一种用于端到端目标检测的简单高效的流程,它通过同时利用密集检测(dense detection)和稀疏集合检测(sparse set detection),Efficient DETR 在初始化 object containers 利用密集先验,并且弥补了 1 层 decoder 结构和 6 层 decoder 结构的性能差异。 Read more »
Deformable DETR Posted on 2021-11-24 | In Object Detection | Visitors: Words count in article: 2,271 | Reading time ≈ 9 Deformable DETR 缓解了 DETR 收敛慢和高计算复杂性的问题。它组合了 deformable 卷积的稀疏空间采样特性和 Transformer 的相关性的建模能力。论文提出的 deformable attention 模块将一小组采样位置作为所有特征图像素中重要的关键元素的预过滤器。 Read more »
复式记账 Beancount 使用 Posted on 2019-07-13 | In Tools | Visitors: Words count in article: 2,823 | Reading time ≈ 13 复式记账 Beancount 使用 Read more »
L2 Normalization Posted on 2019-07-10 | In Deep Learning | Visitors: Words count in article: 871 | Reading time ≈ 3 L2 Normalization Layer Read more »
Pandas Tutorial Posted on 2019-06-23 | In Python | Visitors: Words count in article: 1,839 | Reading time ≈ 11 Pandas Tutorial Read more »
变量——看见社会小趋势 Posted on 2019-06-20 | In Book | Visitors: Words count in article: 1,385 | Reading time ≈ 5 变量——看见社会小趋势 Read more »
High-level Semantic Feature Detection A New Perspective for Pedestrian Detection Posted on 2019-05-29 | In Pedestrian Detection | Visitors: Words count in article: 1,230 | Reading time ≈ 7 CVPR-19 High-level Semantic Feature Detection A New Perspective for Pedestrian Detection Read more »
Deep High-Resolution Representation Learning for Human Pose Estimation Posted on 2019-05-23 | In Deep Learning | Visitors: Words count in article: 1,950 | Reading time ≈ 9 CVPR-2019 Deep High-Resolution Representation Learning for Human Pose Estimation Read more »