Efficient Anchor-Free Unified Framework for Panoptic Part Segmentation in Real-Time and Resource-Constrained Environments

Part segmentation is one of the most important problems in computer vision. In part segmentation objects are simultaneously detected and their parts segmented. While the issue is yet to be resolved, the proposed work introduces a connected, more efficient and anchor-free pathway for object and path segmentation. The proposed solution addresses the prevalent issues of current segmentation models with high computational cost and long inference times. As a result, it is more suitable for real-time applications and resource-constrained settings. The proposed architecture reduces computational complexity and eliminates anchors, a major cost driver in traditional models, to achieve faster segmentation with comparable panoptic and part-level accuracy. Experimental analysis shows that while the method maintains high accuracy compared to other state-of-the-art models and significantly reduces processing time, further evaluation across broader scenarios is needed to fully validate its effectiveness for real-time applications. The proposed architecture, built on a ResNet50 backbone, offers a lightweight design with only 35.6M parameters and a computational cost of 187 GFLOPs, while handling competitive input sizes (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$800\times 1300$ </tex-math></inline-formula>). It achieves strong results on both the Cityscapes Panoptic Parts dataset (PQ: 71.3%, PartPQ: 66.9%) and the Pascal Panoptic Parts dataset (PQ: 64.8%, PartPQ: 55.3%), outperforming existing methods such as Panoptic-PartFormer++ and UPSNet+DeepLabv3+. In general, the contributions include a one-pass segmentation pipeline that is time efficient, higher accuracy, and reduced computational demands, positioning this method as a practical solution for real-world panoptic part segmentation tasks.