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

Hariharan, Bharath, and others. "Hypercolumns in object segmentation and fine-grained localization." In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015. (Citations: 185).

1 Motivation

High-layer feature descriptors emphasize "what" but lack localization information, whereas low-layer feature descriptors focus on "where" yet fall short of being category-sensitive enough. Refer to Figure for details.

2 Pipeline

Refer to Figure, it is employed during the refinement phase. Upscaling is applied to upscale feature maps to the same size.

Then a feature layer is partitioned into an S × S grid (where S equals 5 or 10, depending on the implementation). The logistic regression classifiers are trained for each grid cell. The predicted class label for every spatial position is

the bilinear form of the nearby grid-based prediction functions. The interpolation process is exclusively applied during the testing phase.

3 Implementation Details

Performing classification at every point within a feature map is equivalent to executing a single 1 \times 1 convolution operation. Therefore, for implementing a linear classifier over hypercolumn features, we can divide the processing into segments corresponding to each feature layer. We then apply successive 1 \times 1 convolutions across all feature channels within these segments to generate intermediate activation maps. After obtaining these activation maps, we upscale them proportionally before aggregating them across spatial dimensions.

第4部分 参考文献 [1]. CVPR 2015会议. 链接地址：http://techtalks.tv/talks/hypercolumns-for-object-segmentation-and-fine-grained-localization/61568.