H y p e r R a d a r :

Learning Structurally Consistent Representations for Multi-View Radar Semantic Segmentation

Ali Zia 1†, Muhammad Umer Ramzan 2†, Abdelwahed Khamis 3, Usman Ali 2, Abdul Rehman 2
1School of Computing, Engineering and Mathematical Sciences, La Trobe University, Melbourne, Australia
2School of Engineering and Applied Sciences, GIFT University, Gujranwala, Pakistan
3CSIRO, Australia
† Equal Contribution
Accepted by ECCV 2026
Release Soon Release Soon
Pipeline
Image

Architecture of the proposed HyperRadar framework. Three radar views (RD, RA, AD) are processed by independent encoders and refined using view-specific learnable hypergraphs. (a) Hypergraph refinement: for an extended object like a car with distant reflections spanning multiple range cells, the hypergraph captures these nonlocal dependencies and groups them into a single hyperedge. (b) Unbalanced Optimal Transport (UOT) aligns the refined features across views. The fused tokens are refined via adaptive attention and subsequently decoded into dense segmentation masks. The network is trained end-to-end using joint segmentation and cross-view consistency objectives.

Quantitative

Semantic segmentation performance on the test split of the CARRADA dataset for the RD (Range–Doppler) and RA (Range–Angle) views. From left to right, the columns show the view (RD/RA), method name, intersection-over-union (IoU) scores for the four classes and their mean, followed by Dice scores for the same classes. The best results are shown in bold black, while the second-best results are highlighted in bold blue.

Image
Qualitative
Image

Qualitative comparisons on two test scenes from the CARRADA test split, showing the RGB camera view alongside semantic segmentation outputs from different methods. For each scene, the top row corresponds to the RD (Range–Doppler) view, while the bottom row corresponds to the RA (Range–Angle) view. The columns show: (a) RD/RA inputs, (b) ground truth, (c) our method, (d) TransRadar, (e) TMVA-Net, (f) MVNet, and (g) U-Net. All RD outputs are rotated for visual consistency. Colours indicate semantic classes: blue for Car, green for Cyclist, red for Pedestrian, and black for Background.