How to learn from unlabeled volume data: Self-Supervised 3D Context Feature Learning
Author：Maximilian Blendowski（University of L¨ubeck, Germany）
提出的2.5D的自監督方法，預測不同slice之間某兩點的相對位移。使用兩個預測網絡：直接數值預測和熱力圖方法。主要用於胸片數據。
Scaling and Benchmarking Self-Supervised Visual Representation Learning
Author：Priya Goyal 等（Facebook）
提出了幾個self-supervised 的trick：
-
increasing the size of pre-training data improves the transfer learning performance for both the Jigsaw and Colorization methods
-
Pre-train 的模型越大越好（resnet50>Alexnet）
Self-supervised Spatiotemporal Feature Learning by Video Geometric Transformations
- 處理數據：video
- 方法：
- a set of pre-designed geometric transformations (e.g. rotating 0°, 90°, 180°, and 270°) are applied to each video
- 預測 transformations (e.g. rotations)
Mix-and-Match Tuning for Self-Supervised Semantic Segmentation（2017CVPR）
author: Xiaohang Zhan(港中文)
![]()
- 分爲3steps：1）pre-train learning by colorization；2)M&M tuning； 3）target segmentation task
- M&M tuning：1）對圖像採取圖像塊，去除嚴重重疊的圖像塊，根據標記的圖像真值提取圖像塊對應的 unique class labels（比如車、人） ，將這些圖像塊全部混合在一起。2) fine-tuning the network by triplet loss
Boosting Self-Supervised Learning via Knowledge Transfer
- 整體框架圖
- （a）中的pretext task是Jigsaw++ task
![]()
- In Jigsaw++, we replace a random number of tiles in the grid (up to 2) with (occluding) tiles from another random image
DeepCluster
使用聚類對每個類產生僞標籤，用僞標籤作爲預訓練。