Bottom-Up Foreground-Aware Feature Fusion for Person Search

Abstract

The key to efficient person search is jointly localizing pedestrians and learning discriminative representation for person re-identification (re-ID). Some recently developed task-joint models are built with separate detection and re-ID branches on top of shared region feature extraction networks, where the large receptive field of neurons leads to background information redundancy for the following re-ID task. Our diagnostic analysis indicates the task-joint model suffers from considerable performance drop when the background is replaced or removed. In this work, we propose a subnet to fuse the bounding box features that pooled from multiple ConvNet stages in a bottom-up manner, termed bottom-up fusion (BUF) network. With a few parameters introduced, BUF leverages the multi-level features with different sizes of receptive fields to mitigate the backgroundbias problem. Moreover, the newly introduced segmentation head generates a foreground probability map as guidance for the network to focus on the foreground regions. The resulting foreground attention module (FAM) enhances the foreground features. Extensive experiments on PRW and CUHK-SYSU validate the effectiveness of the proposals. Our Bottom-Up Foreground-Aware Feature Fusion (BUFF) network achieves considerable gains over the state-of-thearts on PRW and competitive performance on CUHK-SYSU.

Inception Convolution and Feature Fusion for Person Search

“…After years of development, the Faster R-CNN [ 16 ] based target detection algorithm can now achieve high performance while also slowing down detection. Some algorithms for pedestrian retrieval [ 17 ] with deeply integrated networks [ 18 ] have achieved reasonable performance. For the one-step approach, they designed a multi-tasking framework [ 19 ] based on Faster R-CNN, establishing a regional proposal network (RPN) to generate region proposal [ 20 ] and then input it into subsequent parallel detection and re-ID branches.…”

“…The first part does not contain residual blocks, which are mainly used for the calculation of convolution, regularization, activation function, and maximum pooling of the input, and the second, third, fourth, and fifth parts of the structure contain residual blocks that do not change the size of the residual blocks but are only used to change the dimensionality of the residual blocks. ResNet-50 is a residual network with lower complexity, more stable performance [ 18 ], and faster convergence compared to VGG 16, and it is suitable for many projects with more accurate results in image classification, target detection, and natural language processing. First, we replaced the general convolution in ResNet-50 with inception convolution in Seq-Net, dynamically enhancing the receptive field of feature diagrams [ 20 , 21 ] without increasing computation or degrading feature diagram resolution.…”