ROI-Gated SAHI: Content-Adaptive Slicing-Based Inference for Efficient Object Detection

Rashid Riyadh; Abd Ullah Khan; Imad Gohar; Muzammil Behzad

cs.CVAug 25, 2026

ROI-Gated SAHI: Content-Adaptive Slicing-Based Inference for Efficient Object Detection

Authors: Rashid Riyadh, Abd Ullah Khan, Imad Gohar, Muzammil Behzad

Organizations: Faculty of Information Technology, City University, Malaysia · Department of Information Convergence Engineering, Kyung Hee University, South Korea · Department of Computer Science, National University of Sciences and Technology, Pakistan · School of Computing and Artificial Intelligence, Sunway University, Malaysia ·

Abstract

Slicing-Aided Hyper Inference (SAHI) improves small object detection in high-resolution images but often spends substantial compute on background tiles. We propose region-of-interest (ROI)-Gated SAHI, an inference-time framework that introduces a lightweight proposer to localize foreground regions and restrict sliced refinement to informative areas. We evaluate the framework in two settings. On the COCO128 full split dataset comprising 128 images, static ROI-gating is slower on average than Full SAHI, achieving a speed ratio of 0.88, and yields a lower [email protected] of 0.6602 compared with 0.7569 for Full SAHI. A simple adaptive routing policy with $τ=$ 0.4 educes the mean latency, achieving a slight gain of 1.02 $\times$ over Full SAHI. On a three-image sparse-to-dense case study, ROI-gating achieves speedups ranging from 0.96 $\times$ to 6.90 $\times$ with a mean speedup of 3.41 $\times$ . These results show that ROI-gating is most beneficial in sparse scenes and requires policy-based routing for robust average behavior.

Figures & tables

Fig. 1 : ROI-Gated SAHI pipeline: (top) baseline full-image patching; (bottom) proposer, ROI merge, coverage check with fallback, and selective refinement. Red dashed box marks injected ROI-gating components.

Fig. 2 : Complete overview of the proposed ROI-Gated SAHI framework, illustrating the sequential flow from Stage A to Stage E. The pipeline begins with high-resolution image input, followed by Stage A (Proposer), where a lightweight model generates coarse ROI candidates. Stage B computes ROI coverage and determines the routing criterion. Based on the threshold condition, the pipeline adaptively switches between Stage C (ROI-Gated SAHI), which selectively processes foreground regions, and Stage D (Full-SAHI), which processes the entire image grid. Finally, Stage E performs global detection fusion by combining outputs from the proposer and refiner branches using NMS to produce the final detections.

Fig. 3 : Principle of ROI-Gated SAHI sliced inference. Left : Baseline SAHI processes the full image grid (blue). Right : ROI-Gated SAHI first identifies a candidate region (green dashed boundary) and then processes only ROI tiles. Both paths detect the same object (red rectangle) in this example, illustrating how selective tiling reduces computation cost in sparse scenes.

D_{p}\leftarrow M_{p}(I)

Algorithm 1 ROI-Gated SAHI

Setting	Images	ROI (%)	Full [email protected]	ROI [email protected]	Full Lat. (ms)	ROI Lat. (ms)	Speed	ROI Faster Count	Agreement F1	Agreement Mean IoU
Overall	128	68.79	0.7569	0.6602	263.73	298.24	0.88	45/128	0.7828	0.9100
Sparse regime ( $R<0.30$ )	18	$<30$	0.6459	0.3630	182.42	154.38	1.18	11/18	0.6981	0.7856
Moderate regime ( $0.30\leq R<0.75$ )	45	30–75	0.7265	0.5605	250.27	285.22	0.88	20/45	0.6727	0.9108
Dense regime ( $R\geq 0.75$ )	65	$>75$	0.8063	0.7712	295.56	347.09	0.85	14/65	0.8824	0.9439
Adaptive policy ( $\tau=0.40$ )	128	dynamic	0.7569	0.7305	263.73	258.46	1.02	routed 26/128	0.9381	0.9638

TABLE I : Quantitative evaluation on COCO128. Speed is reported as mean Full/ROI latency (

>1

favors ROI-Gated). ROI Faster Count reflects per-image gains (Adaptive: routing).

R=\{0.30,0.75\}

\tau=0.40

. F1 and Mean IoU report agreement with Full SAHI.

Fig. 4 : Latency Comparison: average latency for Full SAHI, Static ROI-Gated SAHI, and Adaptive routing (threshold 0.40). This figure compares the average runtime of the three options on the full dataset. Static ROI-Gated is slower overall, while Adaptive routing is the fastest on average because it uses ROI only when ROI area is small.

Fig. 5 : COCO128 full split (128 images): adaptive threshold sweep showing mean adaptive latency and number of images routed to ROI. This figure shows how performance changes when we move the routing threshold. As threshold changes, the number of images sent to ROI changes, and latency follows a curve. The paper setting (0.40) gives near-best latency while keeping routing behavior stable.

Fig. 6 : Thresholding for $\tau$ : adaptive threshold sweep showing mean adaptive latency and number of images routed to ROI. This figure shows how performance changes when we move the routing threshold. As threshold changes, the number of images sent to ROI changes, and latency follows a curve. The paper setting (0.40) gives near-best latency while keeping routing behavior stable.

Fig. 7 : Speed Ratio: per-image speed ratio versus ROI coverage, with break-even and regime boundaries. Each point is one image. Points above 1.0 mean ROI is faster; below 1.0 mean Full SAHI is faster. The plot shows that ROI tends to help more when ROI coverage is low, and tends to lose advantage as ROI coverage increases.

Image	Resolution	Full	ROI	Speed	Speedup	ROI%
		detection	detection	(ms)	( $\times$ )
image1	1200×800	102	72	341/355	0.96	69.9
image2	1536×1024	13	4	579/244	2.38	26.4
image3	1024×1536	2	1	527/76	6.90	2.7

TABLE II : Three-image case-study frame statistics and detection counts.

Metric	Value
Frames Tested	3
Mean Speedup	3.41 $\times$
Mean ROI Coverage	33.0%
Total Full SAHI Detections	117
Total ROI-Gated Detections	77
Detection Agreement (Mean F1)	0.544

TABLE III : Aggregate performance summary for the three-image case study.

Image	Full latency	ROI latency	Speed	ROI
	(ms)	(ms)	( $\times$ )	(%)
img1.jpeg	342	355	0.96	69.9
img2.jpeg	579	244	2.38	26.4
img3.jpeg	526.52	76.25	6.90	2.7

TABLE IV : Per-image latency and ROI coverage.

Fig. 8 : Per-image latency comparison between Full SAHI and ROI-Gated SAHI.

Fig. 9 : (Speedup vs ROI Coverage) Per-image speedup across different ROI regimes.

Image	F1	Precision	Recall	Mean IoU
image1.jpeg	0.49	0.60	0.42	0.89
image2.png	0.47	1.00	0.31	0.87
image3.png	0.67	1.00	0.50	0.98

TABLE V : Agreement of ROI-gated SAHI vs. full-image SAHI (IoU matching).

Panel of Figure 16 — Fig. 10 : Case 1. Dense-scene (69.9% ROI) qualitative comparison showing the Input Image, Proposer Detections, Merged ROI Regions, ROI-Gated SAHI result, and Full SAHI baseline. ROI-gated processing yields near-baseline latency due to limited tile-saving potential.

Panel of Figure 17 — Fig. 11 : Case 2. The moderate-density (26.4% ROI), where selective slicing reduces computation while preserving detection localization quality.

Panel of Figure 18 — Fig. 12 : Case 3. Sparse-scene case (2.7% ROI), where skipping background regions produces the largest efficiency gain.

Panel of Figure 19 — Fig. 13 : Case 2. Dense-scene qualitative comparison showing the Input Image, Proposer Detections, Merged ROI Regions, ROI-Gated SAHI result, and Full SAHI baseline.

Panel of Figure 20 — Fig. 14 : Case 1. Dense-scene qualitative comparison showing the input image, proposer detections, merged ROI regions, ROI-Gated SAHI result, and Full SAHI baseline.

Panel of Figure 21 — Fig. 15 : Case 2. Moderate-scene qualitative comparison showing the Input Image, Proposer Detections, Merged ROI Regions, ROI-Gated SAHI result, and Full SAHI baseline.

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

Adaptive Slicing-Assisted Hyper Inference for Enhanced Small Object Detection in High-Resolution Imagery
Apr 21, 2026Francesco Moretti, Yi Jin, Guiqin MarioRemote Sensing Small Object Detection Small Object Detection
College of Educational Science and Technology, Polytechnic University of Turin
HI-MoE: Hierarchical Instance-Conditioned Mixture-of-Experts for Object Detection
Apr 6, 2026Vadim Vashkelis, Natalia TrukhinaObject Detection Mixture-of-Experts Models
HOIBlender: Blending Lightweight Detection with Vision-Language Priors for Efficient Human-Object Interaction Detection
Sep 20, 2026Junwen Chen, Keiji YanaiEfficient Object Detection Object Detection
Department of Informatics, The University of Electro-Communications, Tokyo, Japan

Apr 21, 2026J/K move · Enter open · S save

Adaptive Slicing-Assisted Hyper Inference for Enhanced Small Object Detection in High-Resolution Imagery

Francesco Moretti, Yi Jin, Guiqin Mario

College of Educational Science and Technology, Polytechnic University of Turin

Remote Sensing Small Object Detection Small Object Detection Aerial Object Detection

Open paper AskPDF arXiv

Deep learning-based object detectors have achieved remarkable success across numerous computer vision applications, yet they continue to struggle with small object detection in high-resolution aerial and satellite imagery, where dense object distributions, variable shooting angles, diminutive target sizes, and substantial inter-class variability pose formidable challenges. Existing slicing strategies that partition high-resolution images into manageable patches have demonstrated promising results for enlarging the effective receptive field of small targets; however, their reliance on fixed slice dimensions introduces significant redundant computation, inflating inference cost and undermining detection speed. In this paper, we propose Adaptive Slicing-Assisted Hyper Inference (ASAHI), a novel slicing framework that shifts the paradigm from prescribing a fixed slice size to adaptively determining the optimal number of slices according to image resolution, thereby substantially mitigating redundant computation while preserving beneficial overlap between adjacent patches. ASAHI integrates three synergistic components: (1)an adaptive resolution-aware slicing algorithm that dynamically generates 6 or 12 overlapping patches based on a learned threshold, (2)a slicing-assisted fine-tuning (SAF) strategy that constructs augmented training data comprising both full-resolution and sliced image patches, and (3)a Cluster-DIoU-NMS (CDN) post-processing module that combines the geometric merging efficiency of Cluster-NMS with the center-distance-aware suppression of DIoU-NMS to achieve robust duplicate elimination in crowded scenes. Extensive experiments on VisDrone2019 and xView, demonstrate that ASAHI achieves state-of-the-art performance with 56.8% on VisDrone2019-DET-val and 22.7% on xView-test, while reducing inference time by 20-25% compared to the baseline SAHI method.

Abstract

Object Detection

Explore similar work

Adaptive Slicing-Assisted Hyper Inference for Enhanced Small Object Detection in High-Resolution Imagery

HI-MoE: Hierarchical Instance-Conditioned Mixture-of-Experts for Object Detection

HOIBlender: Blending Lightweight Detection with Vision-Language Priors for Efficient Human-Object Interaction Detection