Geometric perception on edge must simultaneously achieve two goals that are rarely addressed together: preserve accurate spatial understanding for object localization and operate within stringent latency, memory, energy, and precision constraints. This coupling is particularly evident in oriented bird's-eye-view (BEV) detection, where global context, multi-scale interactions, and geometric fusion affect both prediction quality and hardware dataflow. Conventional deployment flows first compress a model using aggregate quality or complexity proxies and then map the compressed operations to a fixed accelerator interface. This separation risks removing geometric structures that appear redundant under coarse metrics but are critical for precise localization, while leaving hardware-level efficiencies in the retained geometry unexploited. We present FieldRT, an algorithm-hardware co-design methodology that uses a common analysis of geometric sensitivity and data lifetime to derive both the deployment profile and its spatial realization. We deploy FieldRT with Axis-Fusion Network (AFN), a sub-million-parameter oriented-BEV detector, and construct controlled profiles that vary the capacity and organization of its geometric processing. The experiments show that redundancy is task-dependent rather than uniform across metrics. Dense prediction can improve while real instance detection deteriorates; coarse localization may remain stable while strict-IoU precision degrades. Moreover, transformations that are individually safe can become detrimental when combined. FieldRT translates the retained geometric structure into a parameterized streaming architecture that shares field access and storage across spatial operators. The resulting custom-PL primitives achieve production-scale numerical closure on an AMD Kria KV260 and make otherwise infeasible multi-dilation execution realizable.

翻译：暂无翻译