Constraint-Oriented YOLOv7 Optimization for Panoramic Object Detection: Representation, Crowded-Scene Detection, and Deployment-Oriented Model Compression

Panoramic object detection is constrained at three levels of the detection pipeline: equirectangular projection deforms the input representation, crowded scenes weaken feature discrimination and complicate overlapping predictions, and high-resolution imagery increases the computational cost. We organized these challenges in a constraint-oriented framework comprising representation-, detection-, and computation-level branches around a common YOLOv7 reference architecture. The branches were independently configured and evaluated rather than cumulatively optimized as a single detector. At the representation level, spherical rotation augmentation, adaptive sampling, routed attention, and detail-preserving transformation yielded an 86.9% mAP50 and a 62.7% mAP50:95 on full PanoWe, increases of 1.8 and 0.9 percentage points over the branch-specific YOLOv7 reference, while the recall increased by 3.0 points and the precision decreased by 1.1 points. At the detection level, strengthened cross-scale interaction, overlap-aware post-processing, and a normalized Wasserstein distance (NWD) localization term yielded an 86.2% mAP50, a 64.1% mAP50:95, and a 91.1% recall on a manually curated 1264-image crowded-scene subset, with a 1.1-point precision decrease relative to its reference. At the computation level, partial convolution and layer-adaptive magnitude pruning retained an 83.8% mAP50 with 19.4 million parameters and 43.0 theoretical GFLOPs, reductions of 47.4% and 58.9% at a cost of 1.3 mAP50 points. These preserved single-run results support a stage-aligned analysis of the accuracy, recall, and model-complexity trade-offs; they do not establish additive gains across branches or statistically significant superiority. The throughput measured on the recorded hardware platform and an external evaluation on the public 360-Indoor benchmark are additionally reported to characterize the deployment behavior and generalization beyond the in-house collection.