- The paper reviews and compares different one-stage object detection architectures including the YOLO family, SSD-based detectors, and anchor-free methods, analyzing each along speed, accuracy, and hardware feasibility.
- The survey finds that while models like YOLOv7 and RTMDet offer significant speed advantages (up to 300+ FPS), achieving high accuracy (e.g., 55.1 mAP on COCO) often trades off with latency and deployment reliability.
- Deployment adequacy for AV scenarios is largely determined by real-world performance under conditions like small objects, occlusions, and adverse weather, highlighting significant gaps, such as missing detections of safety-critical objects, in current mAP FPS-focused metrics.
Scope and contributions
This paper is a survey and comparative analysis of one-stage object detectors for autonomous driving, not a report on a newly implemented detection system (2608.19014). The authors organize the field along four axes: a chronological review of foundational and modern one-stage architectures, an evaluation of speed–accuracy trade-offs reported in the original publications, an analysis of deployment limitations on resource-constrained edge hardware, and an enumeration of open challenges separating benchmark accuracy from real-world autonomous-vehicle (AV) dependability. The survey covers four sub-families: the YOLO lineage (YOLOv1 through YOLOv10), SSD-based detectors (SSD, DSSD, RefineDet), anchor-free methods (CornerNet, CenterNet, FCOS), and modern dense/efficiency-oriented detectors (RetinaNet, EfficientDet, GFL, VFNet, RTMDet).
A notable methodological caveat stated by the authors: the performance comparison aggregates values reported under different datasets, input resolutions, hardware platforms, and metric conventions. The survey therefore presents these as "survey-style" numbers rather than a controlled benchmark, and marks unreported values as N/R rather than interpolating them.
Taxonomy of one-stage design space
The survey structures the field with two complementary taxonomies. The first classifies all object detection into two-stage (R-CNN family) versus one-stage paradigms, then subdivides one-stage detectors into anchor-based families (default-box matching as in SSD/RetinaNet/EfficientDet; grid-based anchors as in YOLOv2–v7) and anchor-free families (center-based as in YOLOv1, FCOS, YOLOX, YOLOv8–v10; keypoint-based as in CornerNet, CenterNet, ExtremeNet). The second taxonomy groups detectors by deployment-oriented design principle: anchor-free/point-based designs that eliminate anchor hyperparameters, efficiency-optimized designs targeting parameter and FLOP reductions for edge hardware, and NMS-free real-time designs using dual label assignment to remove post-processing bottlenecks.
The technical background identifies four concepts that recur throughout the analysis: bidirectional feature fusion (BiFPN's weighted cross-scale connections in EfficientDet), the decoupling of classification confidence from localization precision (emphasized by ReFPN-FCOS and later formalized by GFL and VFNet), foreground–background class imbalance (addressed by Focal Loss and dynamic loss formulations), and end-to-end detection without NMS (YOLOv10's consistent dual-assignment training). These four threads effectively define the trajectory from early regression-based detection to modern efficiency-aware architectures.
Architectural evolution and its driving constraints
The chronological treatment makes explicit how successive innovations respond to specific failure modes rather than pursuing accuracy in isolation. YOLOv1 established single-pass regression but suffered on small objects and dense scenes; YOLOv2 introduced anchor boxes, batch normalization, and dimension clusters to repair localization; YOLOv3 added multi-scale prediction via Darknet-53 residual backbones; YOLOv4 consolidated CSP connections, Mosaic augmentation, Mish activation, and CIoU loss into a strong anchor-based baseline; YOLOX moved to anchor-free prediction with decoupled heads and SimOTA assignment; YOLOv7 refined backbone-neck-head co-design (E-ELAN); and YOLOv10 removed NMS entirely through consistent dual assignments during training.
On the SSD branch, DSSD addressed SSD's small-object weakness with deconvolutional context at the cost of speed, while RefineDet imported a two-step refinement mechanism into a one-stage pipeline to narrow the accuracy gap with two-stage detectors. Among anchor-free methods, CornerNet's keypoint formulation introduced corner-pooling and embedding-based grouping but proved computationally expensive at inference; CenterNet simplified this to center-point heatmaps with regressed size and offset attributes; FCOS offered fully convolutional per-pixel prediction with centerness weighting. The modern dense family shifted attention to loss design and calibration: RetinaNet's Focal Loss for hard-example mining, GFL's joint classification-quality representation and distributional box regression, VFNet's IoU-Aware Classification Score, EfficientDet's BiFPN plus compound scaling, and RTMDet's balanced backbone-neck design with large-kernel depthwise convolutions.
An implication worth noting for AV practitioners: several of these advances are loss- or assignment-level changes (GFL, VFNet, TOOD-style task alignment) that can be transplanted onto existing backbones, meaning detector improvement is not strictly tied to larger or slower models — a point the survey emphasizes when arguing that optimization choices improve reliability without architectural inflation.
The survey catalogs seven benchmarks relevant to AV detection: KITTI (7,481 train images, 8 classes, historically dominant but small), Waymo Open Dataset (1M+ camera images with LiDAR, mAP/mAPH protocol), nuScenes (1.4M images, 10 classes, full 360° multi-sensor coverage with NDS), BDD100K (100K images emphasizing weather and time-of-day diversity), Cityscapes, Argoverse (3D tracking focus), and COCO (general-purpose pretraining). The authors stress that no single benchmark captures all deployment conditions and that results across datasets should not be treated as interchangeable.
The performance comparison highlights the breadth of the speed–accuracy spectrum:
| Detector |
Dataset |
mAP/AP (%) |
FPS |
Params (M) |
| YOLOv1 |
VOC 2007 |
63.4 |
45 |
N/R |
| YOLOv4 |
COCO |
43.5 |
62 |
64 |
| YOLOX |
COCO |
50.0 |
68.9 |
54 |
| YOLOv7 |
COCO |
51.4 |
161 |
36.9 |
| YOLOv10 |
COCO |
46.3 |
N/R |
7.2 |
| SSD |
VOC 2007 |
72.1 |
58 |
N/R |
| RefineDet |
VOC 2007 |
80.1 |
N/R |
N/R |
| CenterNet |
COCO |
37.4 |
52 |
N/R |
| FCOS |
COCO |
44.7 |
N/R |
N/R |
| RetinaNet |
COCO |
39.1 |
5 |
N/R |
| EfficientDet-D0 |
COCO |
33.8 |
134 |
3.9 |
| GFL |
COCO |
45.0 |
N/R |
N/R |
| VFNet |
COCO |
55.1 |
4.2 |
N/R |
| RTMDet |
COCO |
52.8 |
300+ |
N/R |
Two results stand out numerically. First, RTMDet reports 300+ FPS at 52.8 COCO mAP, and YOLOv7 reaches 161 FPS at 51.4 mAP — both substantially outpacing older anchors like RetinaNet (5 FPS) and CornerNet (4.1 FPS). Second, VFNet achieves the highest COCO mAP in the comparison (55.1%) but at only 4.2 FPS, illustrating the survey's central claim that top accuracy and AV-viable latency remain largely disjoint in the reported literature. The authors caution explicitly that VOC-reported figures (e.g., SSD's 72.1, RefineDet's 80.1) are not comparable to COCO figures due to different protocols and class counts.
Cross-cutting trade-off analysis
The survey's synthesis section draws several conclusions that go beyond cataloging. The primary claim is that no surveyed model simultaneously optimizes accuracy, speed, and deployment feasibility; each occupies a distinct point on the trade-off spectrum. A secondary and somewhat stronger observation is that architectural elegance does not confer deployment advantage: anchor-free designs remove hyperparameter tuning and improve training stability, yet CornerNet's keypoint-grouping overhead makes it less suitable for onboard inference than the anchor-based systems it was meant to supersede. The survey concludes that YOLO-family detectors remain the most deployment-ready architecture precisely because their development has been consistently latency-driven rather than accuracy-driven.
The analysis also widens the notion of "deployment constraint" beyond FPS to include memory, power budgets, embedded compute limits, environmental robustness, and system-integration complexity — using SSD as the illustrative case where highest throughput coexists with unacceptable miss rates on small safety-critical objects such as distant pedestrians and cyclists. The radar-chart comparison presented in the paper is explicitly labeled a qualitative five-point, survey-derived summary, not a new empirical result; the authors are careful not to overclaim benchmark status for it.
Limitations and open challenges
The survey is candid about several weaknesses, both in the field and in its own methodology:
Non-comparable evaluation base: most headline results derive from COCO or Pascal VOC rather than driving-specific data. The authors argue that per-class performance on KITTI/Waymo/nuScenes/BDD100K — particularly for pedestrians, cyclists, and traffic signs under rain, fog, nighttime, and occlusion — is the relevant yardstick for AV readiness, and they call for future comparisons to report AV-specific and adverse-condition results alongside general benchmarks.
Metric insufficiency: mAP and FPS dominate evaluation, but neither measures missed detections of safety-critical objects, degradation under distribution shift, or hardware-level resource usage. The survey flags the absence of deployment-centric metrics as itself an open problem, since current practice can systematically overstate model readiness.
Unresolved technical gaps: small/distant object detection, occlusion handling, adverse-weather robustness, and anchor-free localization failures in dense overlapping scenes remain partially addressed at best. The proposed directions are concrete: stronger multi-scale fusion and small-object-weighted training, coupled image-restoration-plus-detection pipelines for degraded conditions, pruning/quantization/hardware-aware design for edge targets, uncertainty and distance outputs integrated with downstream tracking/planning, and faster keypoint-matching schemes for anchor-free real-time use.
Survey-level caveats: the qualitative multi-dimensional comparisons rest on the authors' interpretation of reported properties rather than controlled experiments, and the aggregated performance table inherits the heterogeneity of its source papers (different input sizes, hardware, and reporting conventions). Readers requiring deployment decisions should treat the table as indicative, not definitive.
Conclusion
This survey consolidates a decade of one-stage detector development into a deployment-oriented framework for autonomous-driving perception. Its principal contribution is connective: it links loss-design innovations (Focal Loss, Varifocal Loss, distributional regression), structural innovations (BiFPN, decoupled heads, NMS-free dual assignment), and anchor-free reformulations to the specific latency, robustness, and hardware constraints of onboard perception. The evidence assembled supports a measured conclusion — one-stage detectors have largely closed the raw accuracy gap while retaining decisive speed advantages, yet the gap between benchmark mAP and dependable real-world AV operation persists across small-object handling, adverse conditions, and evaluation methodology. The open questions the survey leaves are specific and actionable: whether deployment-centric metrics can be standardized, and whether anchor-free and NMS-free designs can match YOLO's real-time profile under driving-domain validation.
