[深度学习论文笔记][Instance Segmentation] Hypercolumns for Object Segmentation and Fine-Grained Localization

Hariharan, Bharath, et al. “Hypercolumns for object segmentation and fine-grained localization.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni-
tion. 2015. (Citations: 185).

1 Motivation

The high layer features care more about “what” but lose localization information, while the low layer features care more about “where” but are not category-level sensitive enough. See Fig.

2 Pipeline

See Fig. This is used in the refinement step. Upsampling is used to resize feature maps to have same size.

Then divide the feature map into S × S grid (S = 5 or S = 10 in our case). A logistic regression classifier is trained for grid cell. The classification prediction of each position is

the bilinear interpolation of nearby grid prediction functions. Interpolations are used only at test time.

3 Implementation Details

Applying a classifier to each location in a feature map is the same as a 1 × 1 convolution. Thus, to run a linear classifier on top of hypercolumn features, we break it into blocks
corresponding to each feature map, run 1 × 1 convolutions on each feature map to produce score maps, upsample all score maps to the target resolution, and sum.

4 References
[1]. CVPR 2015. http://techtalks.tv/talks/hypercolumns-for-object-segmentation-and-fine-grained-localization/61568/.