AdaptIS: Adaptive Instance Selection Network
2019-09-19 12:58:07
Paper: https://arxiv.org/pdf/1909.07829.pdf
Code (MXNet): https://github.com/saic-vul/adaptis
Pretrained model for ToyV1: https://drive.google.com/open?id=1IuJUh0JvbKYILBxCeO2h6U4LG-9DoTHi
Pretrained model for ToyV2: https://drive.google.com/open?id=1RxepfpJF5gRpRNYu1urdV748suF3TL5k
Related Paper:
Panoptic Segmentation, Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, Piotr Dollar
Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization [Paper] [Code] [Blog]
An intriguing failing of convolutional neural networks and the CoordConv solution [Paper] [Code] [Blog]
1. Background and Motivation:
本文提出一种新的分割方式，即：给出一个 BBox，该方法可以将该位置的物体分割出来，而不是全部分割出来。示意图如下所示：
本文所提出方法的名称为：AdaptIS，不依赖于 bounding box proposal。而是直接优化目标分割精度。给定一张图像 I 和 一个固定的 point proposal (x, y)，作者直接优化目标损失函数。我们利用一个 pixel-wise loss 来计算 AdaptIS 预测 和 target object 的 mask。