- Author: Rahul Rama Varior, NTU Singapore; Mrinal Haloi, Nanyang technological Universi; Gang Wang
- Picutre of the model
- Current state-of-the-art on Market-1501.
Contribution
- Architecture of baseline siamese network for person re-id.
- Matching gate between convolutional blocks.
Matching gate (MG) structure
- Feature summarization
- Aggregates the local feature along a horizontal stripe.
- Deal with problem of changed view point (view point change in re-id typically in the horizontal direction, same parts are very likely to be along the same horizontal region).
- Equation and dimention:
;
where is the row of feature map. (maybe ) and (maybe ).
- Feature similarity
- Euclidean distance of and
where - decides the variance of the Gaussian function, learnable, should set a higher initial value.
- Euclidean distance of and
- Filtering and boosting the features
- Repeat c times horizontally to obtain .
- Add filtered feature to original feature.
- Perform L2 normalization across channels after this
Result
- Dataset: Market-1501, CUHK-03, VIPeR.
- Baseline S-CNN outperform most CNN approaches. With MG gaining further improvement.
- Visualization of gate. Low gate activation means low similarity.