Interactive-Segmentation-Papers

2016

[iFCN, CVPR2016] Deep Interactive Object Selection. [Paper]
Key points: first deep-learning-based method, Euclidean distant maps, sampling strategies, refinement by graph cut

[RIS-Net, ICCV2017] Regional Interactive Image Segmentation Networks. [Paper]
Key points: local regional refinement, ROIs sampling, the combination of the global context and local regional segmentation, click discounting factor for training
[BMVC2017] Deep GrabCut for Object Selection. [Paper][Reimplementation in Pytorch]
Key points: transform the loosely-played rectangle as a Euclidean distance map, convolutional encoder-decoder network

[DEXTR, CVPR2018] Deep Extreme Cut: From Extreme Points to Object Segmentation. [Paper][Pytorch]
Key points: exploit extreme points(left-most, right-most, top, bottom pixels), extensive experiemnts on four tasks
[CVPR2018] Interactive Image Segmentation with Latent Diversity. [Paper][Tensorflow]
Key points: multi-modality exploration, plausible segmentations, selection network
[ITIS, BMVC2018] Iteratively Trained Interactive Segmentation. [Paper][Tensoflow]
Key points: iterative training strategy

[BRS, CVPR2019] Interactive Image Segmentation via Backpropagating Refinement Scheme. [Paper][Caffe]
Key points: backpropagating refinement scheme in the test phase to enforce user-specified locations to have correct labels, small adjustments in interaction maps
[CVPR2019] Content-Aware Multi-Level Guidance for Interactive Instance Segmentation. [Paper]
Key points: superpixel-based guidance map rather than distant/gaussian map

[f-BRS, CVPR2020] f-BRS: Rethinking Backpropagating Refinement for Interactive Segmentation. [Paper][Pytorch] Key points: based on BRS, auxiliary variables acting on features are involed, only a small part of network are passed to accelerate
[IOG, CVPR2020] Interactive Object Segmentation with Inside-Outside Guidance. [Paper][Code]
Key points: inside-outside-guidance(click three points), extensive experiments on different datasets and domains
[FCA-Net, CVPR2020] Interactive Image Segmentation With First Click Attention. [Paper][Pytorch]
Key points: first click attention, click loss(weighted binary cross entropy), structural integrity strategy
[arXiv2020] Getting to 99% Accuracy in Interactive Segmentation. [Paper][Pytorch]
Key points: leverage each user interaction, click by click training regime, very high IoU discussions (up to 99%), image/interaction stream design

User interactive ways:
bounding box, clicks(object/background points, extreme points, inside/outside points), contours
Click embedding ways:
distant map, gaussian map, guidance map(combined with superpixel segmentation)
Archietecure:
Encoder: VGG16, ResNet-34/50/101, U-Net, Denset
Decoder: Deeplab-v2/v3+/LargeFOV, U-Net
Post-processing
graph-cuts, CRF, backpropagating refinement, guided filter layer