YOLO-LAN: Versatile YOLO Adaptations

• YOLO-LAN is a versatile, domain-adaptive YOLO paradigm applied for lane detection in driving, polyp detection in endoscopy, and defect detection in optical inspection.
• It employs a single-shot, dense prediction approach that adapts output representations and loss functions to suit diverse application requirements.
• The framework supports real-time inference and is enhanced through optimized loss functions, data augmentations, and the inclusion of negative samples.

YOLO-LAN is an overloaded label in recent arXiv literature rather than a single canonical model name. It appears explicitly as a polyp-detection pipeline in gastrointestinal endoscopy, is used interpretively for a YOLO-based lane-perception formulation in autonomous driving, and is understood in one industrial-inspection paper as a YOLO-based lens anomaly detector. Across these usages, the common substrate is a YOLO-style, single-shot, dense-prediction paradigm adapted to domain-specific targets such as lane markings, road polylines, optical lens defects, or colorectal polyps (Gupta et al., 23 Sep 2025, Wu et al., 2021, Yaseen, 11 Feb 2025, Meyer et al., 2024).

1. Terminological scope and usage

The term has at least three distinct meanings in the supplied literature, and this multiplicity is central to understanding it correctly.

Context	Meaning of “YOLO-LAN”	Primary output
Autonomous driving	YOLO-based lane-perception formulation, especially the lane head in YOLOP and related YOLO-style lane detectors	Binary lane segmentation or line-segment/polyline predictions
Optical inspection	YOLO-based Lens Anomaly/Defect Detection system built on YOLOv8n	Bounding boxes for Lens and Defect
Colonoscopy	“YOLO with Loss, Augmentations, and Negatives”	Bounding boxes for polyps

In the driving literature, the label is not always formalized as a paper title. YOLOP describes “how YOLOP’s lane head operationalizes a YOLO-based approach (YOLO-LAN),” and YOLinO++ states that it is “a concrete instantiation of ‘YOLO-LAN’—a YOLO-style lane detector.” By contrast, the endoscopy paper explicitly titles the method “YOLO-LAN: Precise Polyp Detection via Optimized Loss, Augmentations and Negatives.” The optical-lens paper does not explicitly coin the term, but its details section states that YOLO-LAN can be understood there as a YOLO-based Lens Anomaly/Defect Detection system (Wu et al., 2021, Meyer et al., 2024, Gupta et al., 23 Sep 2025, Yaseen, 11 Feb 2025).

2. Lane-oriented YOLO-LAN in autonomous driving

A lane-oriented interpretation of YOLO-LAN is most directly represented by YOLOP, a panoptic driving perception network that performs traffic object detection, drivable area segmentation, and lane detection simultaneously. YOLOP uses one shared encoder and three decoders. The shared encoder consists of a CSPDarknet backbone and an SPP+FPN neck; the detection branch further uses PAN for bottom-up path aggregation. The lane line head is a semantic segmentation branch identical in architecture to the drivable-area head. It consumes the bottom FPN feature of size $(W/8, H/8, 256)$, applies three nearest-neighbor upsampling stages, and outputs $(W, H, 2)$ logits for “lane line” versus “background.” The representation is explicitly binary per-pixel lane-marking segmentation rather than lane-instance clustering or parametric fitting. On BDD100K validation images resized to $640 \times 384$, YOLOP reports lane Accuracy $70.50\%$, IoU $26.20\%$, and end-to-end throughput of $41$ FPS on an NVIDIA TITAN XP, with $23$ FPS on Jetson TX2. Lane supervision follows Enet-SAD’s comparison protocol: centerlines are computed from BDD100K’s two-line lane annotations, drawn at width $8$ px for training and evaluated at width $2$ px. The lane loss is

$$L_{\mathrm{ll-seg}} = L_{\mathrm{ce}} + L_{\mathrm{IoU}},$$

with

$(W, H, 2)$0

and the full multitask objective is

$(W, H, 2)$1

The paper reports that end-to-end training matches or exceeds step-by-step routines and that the grid-based prediction mechanism aligns naturally with semantic segmentation, mitigating multi-task interference relative to a region-based alternative (Wu et al., 2021).

A more explicitly geometric variant appears in YOLinO++, which follows the “one shot philosophy of YOLO” for detection of 1D structures such as lane boundaries. Instead of producing a binary mask, it overlays the image with a grid and lets each cell host multiple predictors for local line segments. Segment geometry can be represented either by Cartesian endpoints or by midpoint-and-direction,

$(W, H, 2)$2

with reconstruction

$(W, H, 2)$3

Responsibility is assigned either dynamically by Hungarian matching or via anchors derived manually or by k-means clustering. On Argoverse validation, dynamic assignment yields the best retrieval metrics, approximately $(W, H, 2)$4, Recall $(W, H, 2)$5, Precision $(W, H, 2)$6, and Accuracy $(W, H, 2)$7. The method supports crossings, merges, and bidirectional geometries, and uses a specialized line-segment NMS plus polyline chaining to reconstruct higher-level lane structures. In this formulation, YOLO-LAN denotes a YOLO-style lane detector whose geometry is fundamentally line-centric rather than mask-centric (Meyer et al., 2024).

These two lane-oriented usages show that YOLO-LAN is not tied to a single lane representation. In one branch of the literature it denotes full-resolution binary lane segmentation embedded in a multitask scene parser; in another it denotes direct, single-shot estimation of structured 1D lane primitives.

3. YOLO-LAN as “Loss, Augmentations, and Negatives” for polyp detection

In gastrointestinal endoscopy, YOLO-LAN is an explicit acronym for a YOLO-based polyp detection pipeline trained with optimized localization loss, versatile data augmentations, and negative data. The motivation is clinical: the paper states that colorectal cancer accounted for “approximately 1.9 million new cases with more than 935,000 deaths” and that manual colonoscopic miss rates range “from 14% to 30%, depending on the type and size of the polyps.” The method uses standard Ultralytics YOLOv8 and YOLOv12 “without any architectural changes,” fixed input size $(W, H, 2)$8, randomized $(W, H, 2)$9 train/validation/test splits, Adam with initial learning rate $640 \times 384$0, batch size $640 \times 384$1, and early stopping with patience $640 \times 384$2. The five-step pipeline is: add negatives from PolypGen2021 and Kvasir, apply augmentations, split into train/validation/test, replace CIoU with M2IoU, and train/evaluate with precision, recall, F1, mAP$640 \times 384$3, and mAP$640 \times 384$4 (Gupta et al., 23 Sep 2025).

The loss replacement is the paper’s central technical intervention. Let the ground-truth box be $640 \times 384$5 and the prediction be $640 \times 384$6. Define

$640 \times 384$7

$640 \times 384$8

If $640 \times 384$9 is the diagonal length of the smallest enclosing box and $70.50\%$0, then

$70.50\%$1

The paper states that a smaller $70.50\%$2 such as $70.50\%$3 imposes a greater penalty on the corner farther from ground truth, improving stability and convergence. Augmentation is clinically motivated and expands the training set ten-fold. Reported operations include horizontal, vertical, and combined flips; $70.50\%$4 rotation; Gaussian blur with kernel size $70.50\%$5; and a composite blur-plus-rotation transformation. Negative frames are added at $70.50\%$6 and $70.50\%$7 of target training data in order to reduce over-detection bias on polyp-free frames.

On Kvasir-seg, the reported YOLOv8-s configuration with M2IoU, augmentations, and $70.50\%$8 negatives achieves precision $70.50\%$9, recall $26.20\%$0, F1 $26.20\%$1, mAP$26.20\%$2 $26.20\%$3, and mAP$26.20\%$4 $26.20\%$5. The abstract reports, for YOLOv12 on Kvasir-seg, mAP$26.20\%$6 $26.20\%$7 and mAP$26.20\%$8 $26.20\%$9; Table 6 reports best mAP$41$0 $41$1 for YOLOv12n and best mAP$41$2 $41$3 for YOLOv12l under “M2IoU + Augmented data + 20% Negative data.” On BKAI-IGH NeoPolyp, YOLOv8-s under the same setting reports precision $41$4, recall $41$5, mAP$41$6 $41$7, mAP$41$8 $41$9, and F1 $23$0. The Kvasir-seg ablation trajectory for YOLOv8-s is reported as mAP$23$1 $23$2 with CIoU, $23$3 with M2IoU, $23$4 with added augmentations, and $23$5 with $23$6 negatives. The paper characterizes this as a $23$7 gain in mAP$23$8 across the progression, and states that the strongest gains occur under stricter localization thresholds, which is consistent with the intended effect of M2IoU (Gupta et al., 23 Sep 2025).

4. YOLO-LAN in optical lens defect inspection

In industrial inspection, YOLO-LAN refers to a YOLOv8n-based Lens Anomaly/Defect Detection system for mass-produced optical lenses. The task setting is top-down inspection above a conveyor belt using a high-resolution digital camera and controlled lighting to reduce glare and reflections. The model uses Ultralytics YOLOv8n, input resolution $23$9, a C2f-based backbone with SiLU activation, an SPPF neck, and a multi-scale anchor-free detection head employing Distribution Focal Loss for box localization. The detector is trained by transfer learning on a custom dataset with two classes, Lens and Defect; defect examples include scratches, chips, and accumulated debris. Annotations are bounding boxes created in Roboflow, and preprocessing includes auto-orientation and resizing with aspect-ratio-preserving black padding. Reported augmentations include rotation, flips, spatial shifts, Gaussian blur, and scenarios with lighting variation and occlusions (Yaseen, 11 Feb 2025).

The paper does not report total dataset size, split sizes, optimizer, learning-rate schedule, batch size, epochs, regularization, NMS threshold, training hardware, or public dataset or repository links. Its reported evaluation uses $8$0 images containing $8$1 instances. At IoU threshold $8$2, overall performance is precision $8$3, recall $8$4, mAP@0.5 $8$5, and [email protected]:0.95 $8$6. Per class, Lens reaches precision $8$7, recall $8$8, [email protected] $8$9, and [email protected]:0.95 $2$0, whereas Defect reaches precision $2$1, recall $2$2, [email protected] $2$3, and [email protected]:0.95 $2$4. For deployment, the paper reports $2$5 FPS real-time inference on Jetson Nano in the camera-over-conveyor scenario. It also reports preprocessing $2$6 ms, inference $2$7 ms, and postprocessing $2$8 ms per image, totaling approximately $2$9 ms or about $$L_{\mathrm{ll-seg}} = L_{\mathrm{ce}} + L_{\mathrm{IoU}},$$0 FPS; the details section explicitly notes that this discrepancy suggests different hardware contexts and that only Jetson Nano is named for deployment. The paper attributes weak defect performance primarily to annotation limitations and class imbalance, especially incomplete labeling of many tiny defects, which inflates apparent false positives and false negatives (Yaseen, 11 Feb 2025).

This usage differs sharply from the lane and polyp variants. It is a conventional object detector with two classes and no architectural departures from standard YOLOv8n beyond transfer learning, but it retains the same single-shot, edge-deployable logic that motivates other YOLO-LAN interpretations.

5. Shared design patterns across YOLO-LAN variants

Despite the terminological ambiguity, the literature shows several recurring design motifs. First, all variants are single-shot systems that avoid proposal stages. YOLOP uses dense, shared multi-scale features with three heads; the polyp pipeline keeps standard Ultralytics YOLOv8/YOLOv12 heads and alters the localization loss; the lens system uses standard YOLOv8n with an anchor-free head; and YOLinO++ makes the same one-shot commitment for 1D structures by assigning multiple structured hypotheses to each grid cell (Wu et al., 2021, Gupta et al., 23 Sep 2025, Yaseen, 11 Feb 2025, Meyer et al., 2024).

Second, “YOLO-LAN” does not imply a single output representation. YOLOP uses two-channel per-pixel lane logits at full image resolution. YOLinO++ regresses line-segment geometry and reconstructs polylines by chaining adjacent segments. The lens and polyp systems emit bounding boxes. A plausible implication is that the term functions more as a family resemblance—YOLO-derived dense prediction adapted to a specialized target—than as a stable architectural specification.

Third, loss design is highly domain-specific. YOLOP combines focal loss and CIoU for detection with cross-entropy for drivable-area segmentation and cross-entropy plus IoU loss for sparse lane markings. The lens detector explicitly mentions Distribution Focal Loss for localization, but does not report the remaining loss details. The polyp detector replaces CIoU with M2IoU to emphasize corner-level localization precision. YOLinO++ uses a weighted sum of localization, classification, and confidence losses on matched predictors, with dynamic assignment or anchor-based responsibility affecting how supervision is distributed. These differences are not superficial; they track the geometry of the target object, the sparsity of positives, and the deployment regime.

Fourth, each variant uses different strategies to maintain real-time behavior. YOLOP explicitly avoids heavy blocks and uses nearest-neighbor upsampling; the lens system selects YOLOv8n for lightweight edge deployment; the polyp paper emphasizes small YOLO variants and reports strong performance without architectural changes; YOLinO++ uses Darknet-19 and lightweight grid heads, with reported baseline inference around $$L_{\mathrm{ll-seg}} = L_{\mathrm{ce}} + L_{\mathrm{IoU}},$$1 ms per image at batch size $$L_{\mathrm{ll-seg}} = L_{\mathrm{ce}} + L_{\mathrm{IoU}},$$2 for the $$L_{\mathrm{ll-seg}} = L_{\mathrm{ce}} + L_{\mathrm{IoU}},$$3 grid configuration. Across domains, the recurring concern is not only accuracy, but latency under operational constraints.

6. Limitations, misconceptions, and research direction

A common misconception is that YOLO-LAN names a single standardized architecture. The literature does not support that interpretation. In endoscopy, it is an explicit acronym standing for “Loss, Augmentations, and Negatives.” In driving, it can denote a YOLO-based lane subsystem or a YOLO-style lane detector. In optical inspection, it is an interpretive label for a YOLOv8n anomaly detector, and the paper explicitly notes that the authors do not coin the term. Any technical discussion therefore has to identify the domain and paper before the label becomes precise (Gupta et al., 23 Sep 2025, Wu et al., 2021, Yaseen, 11 Feb 2025, Meyer et al., 2024).

Each branch of the literature also reports distinct failure modes. YOLOP’s lane head can struggle under extreme occlusion or heavily faded markings, although joint training reduces false positives near vehicles and non-drivable regions. The lens detector underperforms on defects because of incomplete labeling of small defects and class imbalance. The polyp detector still exhibits overly extended boxes, misses for small polyps masked by overlays, and uncertainty about the optimal ratio of negative frames; the authors recommend temporal modeling, exploration of negative ratios beyond $$L_{\mathrm{ll-seg}} = L_{\mathrm{ce}} + L_{\mathrm{IoU}},$$4–$$L_{\mathrm{ll-seg}} = L_{\mathrm{ce}} + L_{\mathrm{IoU}},$$5, larger multi-center datasets, and prospective clinical trials. YOLinO++ is affected by ground-truth noise, map projection errors, and anchor collisions in complex cells, and explicitly discourages mirroring augmentation because it flips direction-of-travel semantics.

Taken together, these studies indicate that YOLO-LAN is best understood as a domain-adapted YOLO research pattern rather than a singular model family. The persistent themes are geometry-aware supervision, lightweight dense prediction, and task-specific data engineering—whether through multi-task feature sharing for lanes, negative-frame integration for colonoscopy, or annotation refinement for defect inspection. This suggests that future uses of the label will likely remain application-specific and will continue to differentiate themselves primarily through representation choice, loss construction, and dataset design rather than through a universal backbone-head template.