Tile the image, run YOLO per tile, merge the boxes, report the higher small-object recall. That is how most perception teams reach for SAHI, and the number does go up. What most reports omit is that slicing changes the evidence surface — and on a high-integrity perception function, the omission is exactly what a reviewer catches. SAHI stands for Slicing-Aided Hyper Inference. It is an inference-time strategy, not a model change: you keep your trained YOLO detector, but instead of running one forward pass over the full frame, you cut the frame into overlapping tiles, run the detector on each tile, and stitch the per-tile detections back together. The premise is sound. A small or distant object that occupies a handful of pixels in a downscaled full-frame pass occupies many more pixels in a tile, so the detector has more signal to work with. Recall on the small-object bands rises. The trap is treating that recall gain as free. Slicing introduces new detection behaviour — overlap ratios that decide how much context each tile carries, tile-boundary merges where one object straddles two tiles, and non-maximum suppression (NMS) that now has to reconcile detections coming from different tiles rather than one coherent forward pass. Each of those is a decision. Each decision is a place a reviewer will ask, “how did you validate that?” A team that folds only the aggregate recall number into the validation pack gets sent back to explain boundary handling and duplicate merges. A team that traces its slice size, overlap ratio, and merge logic to specific test evidence has already put the rigour where the assigned ASIL level demands it. What problem does slicing actually solve? A single-pass YOLO forward at native resolution has a fixed input size — commonly 640×640 or similar. When your camera delivers a 1920×1080 or higher-resolution frame, the model either downscales the whole frame to its input size or the pipeline crops it. Downscaling is the usual path, and it is where small objects die. An object that spanned 20 pixels in the full frame might survive as 6 or 7 pixels after downscaling — below the effective receptive support the detector needs to fire confidently. The object was in the data; the resolution budget threw it away. Slicing spends the resolution budget differently. Instead of asking the detector to see the whole scene at once, it asks the detector to see one region at a time at closer to native scale. A distant pedestrian, a cyclist at the edge of range, a piece of debris on the carriageway — these are precisely the object-size bands that a full-frame pass under-detects and precisely the bands that carry disproportionate hazard weight in automotive perception. That alignment is why slicing is attractive here and not merely a generic accuracy trick. It is worth being clear about the boundary of the technique. SAHI does not retrain or improve the model’s underlying discrimination — it changes what the model is shown. If your detector cannot recognise a partially occluded cyclist at all, slicing will not teach it to. If the failure is a resolution failure, slicing addresses the resolution. Distinguishing which failure you have is the first diagnostic step, and it is a different question from what a YOLO fine-tune can and cannot transfer when the gap is in the learned representation rather than the input scale. How slice size, overlap, and merge shape what you recover — and what breaks Three parameters carry most of the behaviour, and they interact. Slice size sets how large each tile is relative to the detector’s input. Tiles that are large relative to the input reintroduce the downscaling problem inside each tile; tiles that are small give the detector more effective resolution but multiply the number of forward passes and increase the chance of cutting objects. Overlap ratio controls how much neighbouring tiles share. Overlap exists so that an object falling on a tile boundary appears whole in at least one tile. Too little overlap and boundary objects get split — the detector sees a fragment in each tile and either misses it or emits two low-confidence partial boxes. Too much overlap and you pay for redundant forward passes and generate more duplicate detections to reconcile. The merge step — typically a form of NMS or greedy box matching applied across all tiles after remapping detections to full-frame coordinates — decides what survives. This is where cross-tile behaviour diverges most sharply from single-pass inference. In a single forward pass, NMS reconciles overlapping boxes from one coherent feature map. Across tiles, NMS must reconcile boxes produced by independent passes that saw different context, and it must do so without either dropping a genuine second object nearby or keeping two copies of one object seen in two tiles. Two failure modes follow directly, and they are the ones a reviewer probes first: Duplicate detections. The same object, seen in two overlapping tiles, produces two boxes. If the merge IoU threshold is loose, both survive and the object is double-counted. Double-counting inflates recall and corrupts any downstream tracker that assumes one detection per object. Split objects at tile boundaries. An object bisected by a boundary with insufficient overlap appears as two partial detections, or as none. The merge step cannot reassemble two fragments into one correct box because it has no notion that they belong together — it only knows box overlap, and two non-overlapping fragments do not overlap. The uncomfortable part is that both failure modes can coexist with a higher aggregate recall number. You can raise the headline metric while introducing duplicates in the near field and splits at the edges. That is why the aggregate is not the evidence — the behaviour on the specific size and distance bands is. Quick reference: SAHI parameters and their evidence obligations Parameter / step What it controls Primary failure if wrong Evidence a reviewer expects Slice size Effective resolution per tile Under-detection (too large) or excessive splitting (too small) Recall by object-size bin at the chosen size vs alternatives Overlap ratio Boundary object completeness Split objects (too low), duplicates + latency (too high) Boundary-object recall and duplicate rate as overlap varies Merge IoU threshold Which cross-tile boxes survive Double-counting (loose) or dropped adjacent objects (tight) Duplicate rate and adjacent-object retention on a targeted set Full-frame + slice fusion Whether large objects also detected Large-object misses if slices only Recall across the full size range, not just small bins Confidence handling Per-tile score calibration Boundary boxes suppressed or over-kept Score distribution of merged vs single-pass detections Each row is a claim you are making about the system’s behaviour. On a validation pack, a claim without a test behind it is a liability, not a feature. How does the recall gain map onto ASIL severity? This is where the technique stops being a computer-vision detail and becomes a safety-evidence question. Automotive functions are classified by ASIL (Automotive Safety Integrity Level) according to the severity, exposure, and controllability of the hazards they touch. A function that detects a pedestrian entering the path of a vehicle at speed sits at the high end of that scale; the object-size and distance bands where SAHI helps most — small, distant objects — are frequently the same bands that dominate high-severity hazards, because distance buys reaction time and the objects are small precisely because they are far away. That correspondence is the reason to care about slicing rigorously rather than casually. When the recall you gain lands on a low-severity band, the depth of evidence you owe is modest. When it lands on the band feeding an ASIL D emergency-braking decision, every parameter that produced the gain becomes part of the safety argument. The depth of validation is set by the classification of the function, not by the elegance of the technique — the same graded-rigour logic that governs what an autonomous-vehicle model owes an ASIL D pack applies to the inference strategy layered on top of it. The practical consequence: report recall by size and distance bin, mapped to the ASIL of the function each bin feeds — not a single aggregate number. Aggregate recall hides whether your gain landed where the severity is. A reviewer for a high-integrity function is not asking “did recall go up?” They are asking “did the objects that could kill someone get detected, and can you show me how you know?” Which slicing decisions become evidence surfaces? Every parameter in the table above is an evidence surface once the perception function carries a meaningful ASIL classification. The discipline is to tie each one to a specific test rather than to a general claim of improvement. For slice size and overlap, the test is a recall curve across object-size and distance bins as the parameter varies, so the chosen value is justified by data rather than convention. For the merge step, the tests are two: a duplicate-detection rate on a set constructed to contain objects in overlap regions, and an adjacent-object retention check on a set with genuinely distinct objects close together — because a merge threshold tuned to kill duplicates can also kill real neighbours. For the full-frame-plus-slice fusion decision, the test is recall across the whole size range, proving that shrinking the small-object miss rate did not open a large-object miss rate. This is the same instinct that distinguishes a defensible detector choice from a convenient one, which is why the validation-pack view of choosing between DETR and YOLO and the ASIL-D-constrained comparison of RT-DETR and YOLO both frame the question around what the pack must show, not which architecture wins in the abstract. Slicing is another layer where the same question applies: what test justifies this parameter? In our experience reviewing perception evidence, the clarification rounds that eat the most calendar time are the ones where a metric improved but the mechanism behind the improvement was never documented. Tying each slicing parameter to the test that justified it is not bureaucratic overhead — it is the difference between one review cycle and three. This is an observed pattern across the engagements we have worked on, not a benchmarked figure. When is slicing worth its latency cost? Slicing multiplies inference cost. Running the detector on nine overlapping tiles instead of one frame is, to first order, close to nine forward passes plus the merge overhead. On an embedded automotive compute budget with a hard per-frame latency ceiling, that is not free, and it is not always justified. The trade-off is governed by the function’s ASIL level and by where its hazards live in size-and-distance space. A function whose safety case depends on detecting small distant objects — the bands slicing recovers — has a strong reason to spend the latency, and the ASIL classification may effectively require the recall that only slicing delivers. A function operating on near-field, large objects gains little from slicing and should not pay for it. Between those poles, the decision is empirical: measure the recall gain on the bands that matter for this function, weigh it against the measured latency cost on your target hardware, and let the severity of the function set how much recall you are obliged to buy. There is a middle path worth naming. Slicing does not have to be all-or-nothing per frame. Adaptive strategies — slicing only regions of interest, or only at longer range where the small-object problem concentrates — can recover most of the recall for a fraction of the latency. Whether that complexity is worth carrying into a safety case is itself a judgement, because an adaptive slicing policy is one more decision surface a reviewer will probe. FAQ How does sahi yolo work? SAHI (Slicing-Aided Hyper Inference) keeps your trained YOLO detector unchanged but replaces the single full-frame forward pass with several passes over overlapping tiles of the image, then merges the per-tile detections back into full-frame coordinates. In practice it trades inference latency for higher small-object recall, and it introduces new decisions — overlap, tile-boundary handling, cross-tile NMS — that must be validated rather than assumed. What problem does SAHI’s slicing-aided inference solve that a single-pass YOLO forward at native resolution does not? A single full-frame pass downscales high-resolution camera frames to the detector’s fixed input size, shrinking small and distant objects below the effective resolution the model needs to fire. Slicing spends the resolution budget region by region, so those objects appear at closer to native scale and are recovered. It addresses resolution loss specifically; it does not improve the model’s underlying discrimination. How do slice size, overlap ratio, and the tile-merge / NMS step affect which objects SAHI recovers and where new errors appear? Slice size sets effective per-tile resolution; overlap ratio controls whether boundary objects appear whole in at least one tile; the merge/NMS step reconciles boxes from independent tiles. Too little overlap splits objects at boundaries, too much creates duplicates and latency, and a loose merge threshold double-counts while a tight one drops genuinely adjacent objects. How does improved small-object recall from SAHI map onto the object-size and distance bands that carry the highest ASIL severity? The small, distant object bands that slicing recovers best are frequently the same bands that dominate high-severity automotive hazards, because distance and small size go together. That is why recall should be reported by size and distance bin mapped to the ASIL of the function each bin feeds, not as a single aggregate — the aggregate hides whether the gain landed where the severity is. Which SAHI parameters and merge decisions become evidence surfaces a reviewer will probe, and how do we tie each to a specific test? Slice size, overlap ratio, merge IoU threshold, full-frame/slice fusion, and confidence handling all become evidence surfaces once the function carries a meaningful ASIL level. Each ties to a specific test: recall curves across bins for slice size and overlap, a duplicate rate plus adjacent-object retention check for the merge threshold, and full-size-range recall for the fusion decision. What are the failure modes of slicing — duplicate detections, split objects at tile boundaries — and how do we validate that they are handled? Duplicate detections arise when one object seen in two overlapping tiles survives twice through a loose merge; split objects arise when insufficient overlap bisects an object into fragments the merge step cannot reassemble. Validate duplicates with a set built to contain objects in overlap regions, and validate splits with boundary-object recall as overlap varies — both alongside the aggregate, never in place of it. When is slicing-aided inference worth its added latency cost for a given perception function, and how does the ASIL level inform that trade-off? Slicing costs roughly one forward pass per tile plus merge overhead, so it is justified when the function’s hazards live in the small/distant bands slicing recovers and the ASIL level effectively requires that recall. For near-field, large-object functions the gain is small and the latency is not worth it; between those poles, measure the recall gain and latency on target hardware and let the function’s severity set how much recall you must buy. If you are building this into a camera-based perception stack, the open question is rarely “does slicing help?” — it usually does. The harder question is whether your validation for computer vision systems treats each slicing parameter as an evidence surface graded to the function’s ASIL, or as a switch that quietly raised a number. The reviewer will assume the latter until you show the former.
