DetOH: An Anchor-Free Object Detector with Only Heatmaps
R Wu, X Xiao, G Hu, H Zhao, H Zhang… - … Conference on Advanced …, 2023 - Springer
R Wu, X Xiao, G Hu, H Zhao, H Zhang, Y Peng
International Conference on Advanced Data Mining and Applications, 2023•SpringerIn object detection, the anchor-based method relies on too much manual design, and the
training and prediction process is too inefficient. In recent years, one-stage anchor-free
methods such as Fully Convolutional One-stage Object Detector (FCOS) and CenterNet
have made a splash in object detection. They not only have a simple structural design, but
also demonstrate competitive performance. They have exceeded many two-stage or anchor-
based approaches. However, in industrial applications, the design of its multiple output …
training and prediction process is too inefficient. In recent years, one-stage anchor-free
methods such as Fully Convolutional One-stage Object Detector (FCOS) and CenterNet
have made a splash in object detection. They not only have a simple structural design, but
also demonstrate competitive performance. They have exceeded many two-stage or anchor-
based approaches. However, in industrial applications, the design of its multiple output …
Abstract
In object detection, the anchor-based method relies on too much manual design, and the training and prediction process is too inefficient. In recent years, one-stage anchor-free methods such as Fully Convolutional One-stage Object Detector (FCOS) and CenterNet have made a splash in object detection. They not only have a simple structural design, but also demonstrate competitive performance. They have exceeded many two-stage or anchor-based approaches. However, in industrial applications, the design of its multiple output heads hinders the installation of the model. At the same time, different output heads mean a combination of multiple loss functions. This introduces problems in training. Here, we propose an anchor-free object Detector with Only Heatmaps (DetOH) to solve object detection. Bounding box parameters are calculated by post-processing. The design of the single output head allows object detection to use a semantic segmentation network, realizing the unification of the two frameworks. In addition, compared with CenterNet, we have greatly improved the speed of object detection (6 vs. 32 Frames Per Second) with 3.2% Average Precision boost. The proposed DetOH framework can be applied to multi-target tracking, key point detection and other tasks.
Springer
Showing the best result for this search. See all results