Transfer learning: Contains all source or target, labeled or unlabeled, finetune, multi-task learning situations. In a word, everything except the use of ( a single data set for supervised learning ) during training is counted in transfer learning.
supervised learning: training data has labels (single data set)
unsupervised learning: Unlabeled training data (single data set)
semi supervised learning: part of the training data is labeled and part of the training data is unlabeled (single data set)
doman adaptation: multiple data sets, target has no label, source has label (multiple data sets)
Weakly supervised learning: There are labeling errors, or incomplete labeling, or segmentation tasks only need to be detected labels, etc. (single data set)
See the figure below for more explanation:
Summary of settings appearing in the blog post:
Setting1: universal semi-supervised semantic segmentation. Both the target and the source have a label, and some do not have a label.
Setting2: cross-category semi-supervised semantic segmentation, in-category semi-supervised semantic segementation。
Setting3: partial domain adaptation, both source and target have tags, and there are more source tags than target.
Setting4: Beyond domain adaptation, source domain has image & label, target domain cannot be seen during training, only test can see.
The following articles are mainly Segmentation
Some articles are also doing classification and detection tasks
1. Weakly- and Semi-Supervised Panoptic Segmentation(ECCV, 2018)：
Setting: weakly- and semi-supervised semantic and instance segmentation. Weakly: Use the detection box and category labels to train to obtain the image segmentation results. semi: add a part of human-annotations (can be understood as: the segmentation label of the human class)
Method: Single data set, combining the results of semantic seg and detector into the network of instance seg, you can get the result of instance seg.
The method to get the result according to the detection frame and object label: train a multi-class network, take the channel of the conv layer for different branches, frame the location of the object label according to the detection frame, and get a rough prediction segmentation result according to the heatmap of the specific location.
Contribution: It is the first time to use weakly supervised non-overlapping instance segmentation. For the first time, semantic and instance segmentation were performed simultaneously in the weakly supervise task. It is the first time to perform weakly- and semi-supervised tasks in cityscapes.
Backbone: psp+resnet101
Datasets: cityscapes, pascal voc. (coco assist)
2. Penalizing Top Performers: Conservative Loss for Semantic Segmentation Adaptation(ECCV, 2018)：
Setting: unsupervised domain adaptation。
Method: Proposed a new loss: conservative loss. Prevent the model from learning too poorly and lack of discrimination, and don't learn too well and fall into over fitting of the source domain. The middle part should be as small as possible, and the two ends should be as large as possible. Those with a probability close to 1 give lower loss to prevent better learning. The probability is close to 0, indicating that this part needs to increase training, give high loss, let it modify the weight more.
Contribution: new loss
Backbone: FCN (vgg16)
Datasets: cityscapes, gtav, synthia
3. Domain Transfer Through Deep Activation Matching(ECCV, 2018)：
Setting: unsupervised domain adaptation。
Method: Add a lot of gan in the middle of the network. The loss of gan is calculated according to jsd (Jensen-Shannon divergence).
Contribution: new loss
Backbone: classification: le-net. segmentation: erf-net
Datasets: classification: mnist, usps, svhn. segmentation: cityscapes, gtav, synthia
4. Partial Adversarial Domain Adaptation (ECCV, 2018): (Follow-up work of this work - Partial transfer learning with selective adversarial networks. )
Setting: partial domain adaptation, source and target have tags, source tags are more than target.
Method: Use adversarial learning in partial domain adaptation.
Contribution: Use adversarial learning in partial domain adaptation.
Backbone: - (on)
Datasets: office-31, office-home, imagenet-caltech, visda2017
4. Conditional Adversarial Domain Adaptation(NIPS, 2018)：
Setting: domain adaptation. Single task, multiple data sets. The target does not have a label, and the source has a label.
Method: The result of multiplying the features extracted in the network and the predicted category is used for confrontation learning.
Contribution：同method
Backbone: alexnet, resnet-50
Datasets: office-31, imageclef-da, office-home
5. Adversarial Multiple Source Domain Adaptation(NIPS, 2018)：
Setting: domain adaptation. Single task, multiple data sets. The target does not have a label, and the source has a label.
Method: Supervised training labeled data, multiple discriminators distinguish each source domain and target domain.
Contribution：同method
Backbone: --
Datasets: mnist, mnist-m, svhn, synthdigits
6. Importance Wegihted Adversarial Nets for Partial Domain Adaptation(CVPR, 2018)：
Setting: partial domain adaptation, source and target have tags, source tags are more than target.
Method: Weighted discriminator (use the first d to calculate the weight, and the second d to reverse the modification coefficient).
Contribution: Discriminator for calculating weights
Backbone: resnet18
Datasets: office+caltech-10 object datasets, office-31 dataset, caltech-256
7. Object Region Mining with Adversarial Erasing: A Simple Classification to Semantic Segmentation Approach(CVPR, 2017)：
Setting: unsupervised semantic segmentation. There are classification labels, and training results are obtained.
Method: For object classification, different areas of the object have different contributions to the classification. There will be a heat map. Remove the high response part (in fact, use the image average instead), continue training, get the heat map, and then iterate in this way to get multiple networks. There are multiple heat maps, which together can roughly separate objects.
Contribution: classification --> segmentation. psl: erasure method
Backbone: deeplab(vgg16)
Datasets: pascal voc 2012
8. Universal Semi-Supervised Semantic Segmentation(arxiv, 2018)：
Setting: universal semi-supervised semantic segmentation. Both the target and the source have a label, and some do not have a label.
Method: A total of two data sets, each data set. In each data set, the data with label is trained by supervised learning method, and the data without label is trained by unsupervised learning method. Unsupervised learning calculates loss based on similarity.
Contribution: New setting, unsupervised learning calculates loss based on similarity
Backbone: resnet18
Datasets: cityscapes, camvid, id, sun rgb-d
9. Transferable Semi-supervised Semantic Segmentation(arxiv, 2018)：
Setting: cross-category semi-supervised semantic segmentation, in-category semi-supervised semantic segementation。
Method: l-net: Combine classification label and segmentation label training to get prediction. p-net: adversarial learning
Contribution：l-net and p-net？？？
Backbone: DeepLab(vgg16) - L-net. CAM(vgg16) - classification net. 6-3*3 conv layer + 3 fully connected layers - P-net.
Datasets: PASCAL VOC 2012.
10. Virtual-to-Real: Leaning to Control in Visual Semantic Segmentation(IJCAI, 2018)：
Can't understand it, it seems to be related to robot? It may not have much to do with us. . Let me record it first. .
Network: A3C (compared to DR-A3C, ResNet-A3C)
Datasets: ADE20K
11. Label Efficient Learning of Transferable Representations across Domains and Tasks(NIPS 2017)：
Setting: labeled source data and unlabeled or sparsely labeled data in the target domain.
Method: 1 supervised loss(source domain & target domain labeled data)+2 GAN loss(source domain feature & target domain feature)+3 Entropy Loss(source feature & unlabeled target feature; labeled target feature & unlabeled target feature )
Contribution：loss3
Backbone: fcn8s + vgg16, drn-26
Datasets: SVHN, MNIST; Image net , ucf-101
12. Beyond Domain Adaptation: Unseen Domain Encapsulation via Universal Non-volume Preserving Models(arxiv, 2018)：
Setting: source domain has image & label, target domain cannot be seen during training, only test can see.
Method: Assume that each domain has a Gaussian distribution, but their mean and variance are different. By changing the mean and variance, data of different domains are generated and added to the training process.
Contribution：同method
Backbone: conv-pool-conv-pool-fc-fc-softmax, real nvp
Datasets: mnist, usps, svhn and mnist-m; yale-b, cmu-pie, cmu-mpie
~~~~~~~~~Updated from time to time~~~~~~~~~
ICLR 2018 100 papers: http://search.iclr2018.smerity.com
CVPR 2018 paper list: http://openaccess.thecvf.com/CVPR2018.py
ECCV 2018 paper list: http://openaccess.thecvf.com/content_ECCV_2018/html/
NIPS 2018 paper list: http://papers.nips.cc