Keyword [Locally Adaptive Learning Loss]
Guo J, Ren P, Gu A, et al. Locally Adaptive Learning Loss for Semantic Image Segmentation[J]. arXiv preprint arXiv:1802.08290, 2018.
1. Overview
1.1. Motivation
- Most loss layer focus on pixel-wise, ignore spatial layout and interaction with neighbouring pixel (not sensitive to intra-class connection)
In this paper, it proposed Locally Adaptive Learning Loss for segmentation
- merge predicted distribution over a small group of neighbouring pixels with same category (imporve the capability of discriminating targets from both inter- and intra- class)
- sliding window + ensemble by Minkowski pooling (focus on high loss, rebalancing)
1.2. Related Work
- contrastive loss, triplet loss and center loss
- pixel-wise loss. collapse the spatial dimension
- Aligned RoI. sigmoid + binary loss (ROI maintain spatial layout)
Loss Max-Pooling. handle imbalanced inter-class dataset in segmentation (assign weight to each pixel based on their losses)
weighted ensemble entropy estimator. better accuracy and higher converge rate
1.3. Locally Adaptive Loss
Selective Ensemble
m. number of the pixel which have the same label
- μ. select same label pixel
- wd. Gaussian weight based on chessboard distance
- x. pixel
when ε is softmax cross-entropy
- Batch Pooling
Mp. number of batch
Minkowski Pooling. when k increases, it will focus on high loss value and increase the impacts of mispredited samples of intra-class
1.4. Dataset & Metric
VOC2012, IoU
1.5. Experiments
- DeepLabV2 (disable multi-scale and CRF)
- Batch size 2
- Crop size 321x321
