Towards Unsupervised Crowd Counting via Regression-Detection Bi-knowledge Transfer

Abstract

Unsupervised crowd counting is a challenging yet not largely explored task. In this paper, we explore it in a transfer learning setting where we learn to detect and count persons in an unlabeled target set by transferring bi-knowledge learnt from regression-and detection-based models in a labeled source set. The dual source knowledge of the two models is heterogeneous and complementary as they capture different modalities of the crowd distribution. We formulate the mutual transformations between the outputs of regression-and detection-based models as two scene-agnostic transformers which enable knowledge distillation between the two models. Given the regression-and detection-based models and their mutual transformers learnt in the source, we introduce an iterative self-supervised learning scheme with regression-detection bi-knowledge transfer in the target. Extensive experiments on standard crowd counting benchmarks, ShanghaiTech, UCF_CC_50, and UCF_QNRF demonstrate a substantial improvement of our method over other state-of-the-arts in the transfer learning setting. CCS CONCEPTS • Information systems → Multimedia information systems; • Humancentered computing → Collaborative and social computing; Visualization; • Computing methodologies → Computer vision.

Dynamic Momentum Adaptation for Zero-Shot Cross-Domain Crowd Counting

“…In [30], multiple domain specific modules are trained using labeled data from target domains for online switching. [27] proposes to take advantage of both detection and regression-based counting frameworks, and fine-tunes the offline learned counter via online-estimated pseudo labels.…”

“…Source domain training on ShanghaiTech. Following the experiment design of other domain adaptation methods [10,27], we first evaluate cross-domain counting using SHA or SHB as the source domain, and SHB/SHA and UCF-QNRF as the target domains. We compare our proposed C 2 MoT with two main types of counting methods: 1) fine-tuning based domain adaptation methods, including Cycle GAN [51], SE CycleGAN [45], SE+FD [10] and RBT [27];…”

Fine-grained Domain Adaptive Crowd Counting via Point-derived Segmentation

“…Hossain et al [21] reduce the domain shift by minimizing the feature distances (i.e., Maximum Mean Discrepancy (MMD) [44]) across domains. (iii) others [41], [77]: [41] introduce an extra head detector for mutual training with the crowd counter. [77] present a neuron linear transformation to optimize a small amount of parameters based on few targetdomain training samples.…”

“…[77] present a neuron linear transformation to optimize a small amount of parameters based on few targetdomain training samples. Methods in others can be regarded as supplements of pixel-level and feature-level methods with additional bounding box annotations [41] or extra targetdomain annotations [77].…”

Self-supervised Domain Adaptation in Crowd Counting

“…source domain, and perform counting on more open-set scenarios, i.e. target domain, [12,13] by transfer learning and domain adaptation techniques. Liu et al [13] enable knowledge distillation between both regression-based and detectionbased models by formulating the mutual transformation of outputs.…”