VALSE2020 重要主题年度进展回顾 (APR)
NAS
其实只是简要介绍了今年综述的论文，笔记在这里。
视觉目标检测
anchor-free object detection
- CornerNet ECCV2018
- CerterNet huawei ICCV2019 center pooling 特征与物体的对应
RepPoints ICCV2019 bounding box 作为特征点的凸包络，回归时满足特征点组成的 box 与 ground truth 损失越来越小
- FCOS ICCV2019 引入 center-ness loss，“变了一种形式但是是有益的探索”
- Feature selective anchor-free module for single-shot object detection CVPR2019
learning-to-match object detection
- FreeAnchor: learning to match anchors for object detection 某些物体如拖把是几何偏心的物体，不用IoU将特征与物体进行匹配。
- autoassign: differentiable label assignment for dense object detection 孙剑
- probabilistic anchor assignment with IoU prediction for object detection
feature decoulping
- revisiting the sibling head in object detector 分别处理 localization 和 classification
NAS
- EfficientDet CVPR2020
DERT
detection with transformer
efficency-based object detection
- YOLO-V4
domain adaptive detection
- strong-weak distribution alignment for adaptive object detection CVPR2019
few-shot object detection
- few-shot object detection via feature reweighting ICCV2019
- frestratingly simple few-shot object detection ICML2020
long-tailed distribution
- LVIS 数据集 cvpr2020
low-resolution objects
- scale match for tiny person detection WACV2020
dense objects
- Precise detection in densely packed scenes cvpr2019
- dynamic refinement network for oriented and densely packed object detection
- NWPU-Crowd 数据集 PAMI2020
模型可解释性
https://github.com/clovaai/wsolevaluation
dataset bias vs model bias
visualing the internal representations
interpretability of self-attention modules
GAN 网络可解释性
GAN dissection
InterFaceGAN
steerability of GAN
- unsupervised discovery of interpretable directions in GANs latent space ICML2020
- closed-form factorization of latent semantics in GANs arXiv 2007.06600
