Visual recognition models typically process a fixed view under a fixed computation budget, even though some images benefit from more detail or more spatial evidence. Active vision can adapt by first observing a coarse scene and then selecting informative fixations, but the public AdaptiveNN-DeiT-S pipeline still fixes both the field of view and the number of observations. Researchers have now extended this fixed-step recognizer with a controllable observation interface that adapts both of these to the input.

The model predicts a per-fixation scale that sets its field of view, refines each selected crop with optional scale-gated foveation and a residual high-resolution detail-token branch, and uses label-grounded stop heads for input-dependent early termination. This allows the same checkpoint to expose a controllable observation budget, so users can trade accuracy for compute by setting a threshold without retraining for each operating point.

Across fine-grained benchmarks, full ImageNet-1K, and ImageNet-A/R-derived splits, the interface remains competitive with fixed-step references. On sequential ImageNet-1K, the learned stop rule outperforms fixed-step truncation of the same checkpoint, sustaining 81.9% top-1 accuracy using 2.45 of the four focus steps and exceeding three-step truncation at lower cost. On each benchmark, a variant combining scale control and detail-token fusion achieves the highest accuracy among the tested non-exit variants.