CABiNet (ICRA 2021) vs YOLO26-sem on UAVid: accuracy, compute, and GPU latency [P]

Disclosure up front : I'm the original first author of CABiNet (ICRA 2021), so I'm not a neutral party. Everything below is reproducible from the repo. Background CABiNet is a dual-branch CNN for real-time semantic segmentation: a high-res spatial branch, a lightweight context branch (global aggregation + local distribution) over a MobileNetV3 backbone, fused with a small FFM. Published 2021, then it went quiet. I came back this year, rebuilt the repo (PyTorch 2.x, Hydra, AMP, EMA, poly-LR, OHEM loss, CI + tests), and used it to ask one question on **UAVid**, the aerial dataset the original paper targeted: how does a purpose-built 2021 efficient architecture compare to a 2026 general multi-task model with a dedicated semantic-segmentation variant? What's actually controlled (and what isn't) Both models run off the same converted dataset and splits , the same ENet inverse-log class weighting (`cls_pw=0.5`), EMA weights for eval , and the same evaluation protocol : single-scale, no test-time augmentation. What is not matched: | Axis | CABiNet | YOLO26-sem | Potential advantage | | --- | --- | --- | --- | | Initialization | ImageNet-pretrained MobileNetV3 backbone; seg layers random | full net pretrained on Cityscapes + ADE20K | potentially favors YOLO | | Epoch budget | 5000 (early stop, patience 100) | 500 (early stop, patience 50) | potentially favors CABiNet | | Optimizer / schedule | SGD + poly decay, decoder LR ×10 | SGD + cosine | different | | Loss | OHEM-CE + aux deep supervision | CE + Dice + aux | different | | Extra augmentation | none | mosaic 0.8, copy-paste 0.15 | potentially favors YOLO | So this is not an architecture-only ablation. It's a controlled benchmark: the data representation, class weighting and evaluation are standardized, while each model keeps a model-specific training recipe. None of the rows above is an isolated experiment, so I haven't measured how much any single one is worth. Results — UAVid test split, 1024×1024, single-scale | Model