In Defense of Single-column Networks for Crowd Counting

CNN-based Density Estimation and Crowd Counting: A Survey

“…Moreover, it can accelerate convergence speed. Therefore, it is used to capture and fuse multi-scale features in SCNet [185], PaDNet [79] and CAN [86] for crowd counting.…”

“…Albeit significant performance improvement has been achieved to a great extent by applying CNNs into density Pandensity /subregion MCNN [1] CSRNet [12] DRSAN [71] DecideNet [7] SCNet [185] PaDNet [79] SAAN [8] PACNN [85] CAN&ECAN [86] SFANet [89] W-Net [90] DSNet [163] L2SM [98] DSSINet [178] SPANet [193] MBTTBF-SCFB [194] BL [192] S-DCNet [195] PGCNet [133] map estimation crowd counting models, there are remain some challenges to be conquered. A robust network should have the capability of coping with various complex scenarios.…”

“…Current CNN-based deep models are designed with a sophisticated structure, which always comes with millions of parameters and the cost of a tremendous increase in computation (FLOPs). Although a great effort has been devoted to making the model efficient, such as CSRNet [12] and SCNet [185], they usually adopt VGG16 [205] or ResNet [208] pre-trained on Imagenet [212], which is a large dataset for classification, but the task of object counting belongs to a regression task. Thus it may affect the performance to some extent.…”

PSGCNet: A Pyramidal Scale and Global Context Guided Network for Dense Object Counting in Remote Sensing Images

“…DSNet [52] cascades multiple dense dilated convolution blocks and link them with dense residual connections. SCNet [53] leverages nested dilation convolutional layer, which incorporates kernels with different dilation rates, and SPP layer with different stride. ADSCNet [54] adopts adaptive dilated convolution to learn dynamic and continuous dilated rates for each pixel location.…”

Enhancing and Dissecting Crowd Counting By Synthetic Data

“…A lot of work has been proposed to improve the performance of detection algorithms. These studies either focus on proposing more advanced network structures (for example multi-column network [1,2], scale aggregation module [3,4] and scale adaptive module [5,6,7]), or focus on designing more suitable loss functions [8]. These two focus points have greatly improved the performance of existing algorithms.…”