Represent, Compare, and Learn: A Similarity-Aware Framework for Class-Agnostic Counting

Accurate few-shot object counting with Hough matching feature enhancement

“…In previous methods (Shi et al, 2022 ; You et al, 2023 ), the exemplar features from multiple shots are usually used to calculate correlation tensor independently, and the commonality of multiple exemplar features is underutilized. Therefore, we build an exemplar feature aggregation (EFA) module, which leverages the features from every exemplar to enhance the commonality.…”

“…It is difficult to obtain the position directly. Previous method (Shi et al, 2022 ; You et al, 2023 ) use adaptive Gaussian kernel to generate the ground-truth density map, but it is difficult to solve the object distortion caused by perspective effect. Here, we use a gaussian smoothing with a fixed size of 16 and a standard deviation of 3.5 to generate the ground-truth density map D gt .…”

“…Here, we use a gaussian smoothing with a fixed size of 16 and a standard deviation of 3.5 to generate the ground-truth density map D gt . Following previous methods (Lu et al, 2018 ; Shi et al, 2022 ; You et al, 2023 ), we use the mean squared error (MSE) loss function.…”

“…In this part, we cascade two times for further feature enhancement. Following prior works (Lu et al, 2018;Shi et al, 2022;You et al, 2023), the counter is composed of several convolution layers and bilinear upsampling layers. The composition of counter is presented in Table 1, and each convolution layer is activated with the Leaky ReLU function.…”

“…However, the query feature is not fully utilized, and the boundary of the density map is fuzzy. Shi et al ( 2022 ) uses self-attention to narrow the distance between the samples and the target objects and then concatenate the similarity map to regress head. But when the objects are dense and the light is dim, the boundary of the density map is not clear, and the accuracy of the count value is low.…”