YOLOv8-D: Continual-Learning Detector for VRU
- YOLOv8-D is a YOLOv8x-based continual-learning detector that employs a switch from Adam to SGD to adapt to new traffic domains.
- It targets vulnerable road users by sequentially training on distinct datasets like VisDrone and Caltech to mitigate catastrophic forgetting.
- The model contrasts with other dynamic YOLOv8 variants by focusing on optimizer-induced adaptability rather than dynamic architectural modifications.
Searching arXiv for the cited YOLOv8-Dynamic and closely related dynamic/adaptive YOLOv8 papers. arXiv search query: "YOLOv8-Dynamic YOLOv8-D continual learning vulnerable road users (Nasir et al., 15 Jul 2025)" YOLOv8-Dynamic (YOLOv8-D) most precisely denotes a continual-learning extension of YOLOv8x for real-time detection of vulnerable road users in complex traffic scenarios. In narrower technical usage, the term refers to the model explicitly named “YOLOv8-Dynamic (YOLOv8-D)” in a single paper; in broader informal usage, it can also refer to several YOLOv8-derived detectors that introduce input-conditioned gating, dynamic convolution, adaptive heads, or dynamic feature fusion, even when those works use different official names and different mechanisms of “dynamic” behavior (Nasir et al., 15 Jul 2025).
1. Terminology and scope
The literal model name YOLOv8-Dynamic (YOLOv8-D) appears in the vulnerable-road-user detection study built from YOLOv8x and framed around continual learning across statistically different traffic datasets. In that paper, “dynamic” primarily denotes temporal adaptability during sequential domain adaptation rather than a dynamically reconfigured backbone, head, or routing graph. By contrast, several other YOLOv8-based papers are often relevant to the same phrase only in a functional or descriptive sense: they add dynamic gating, dynamic convolution, dynamic heads, or dynamic fusion, but they do not formally define their models as YOLOv8-Dynamic or YOLOv8-D (Nasir et al., 15 Jul 2025, Huang et al., 26 Jan 2026, Ibrahim et al., 2 Apr 2025, Liu et al., 2023, Li et al., 2024, Jahin et al., 5 Aug 2025, Xing et al., 2024).
| Designation | Official paper name | Relation to “YOLOv8-Dynamic” |
|---|---|---|
| YOLOv8-D / YOLOv8-Dynamic | YOLOv8-Dynamic | Exact model name; continual-learning extension of YOLOv8x |
| YOLO-DS | YOLO-DS | YOLOv8-based adaptive C2f gating; not named YOLOv8-D |
| Enhanced YOLOv8 traffic-sign model | Enhanced YOLOv8 | YOLOv8 with ODConv and other adaptive modules; not named YOLOv8-D |
| ADA-YOLO | ADA-YOLO | YOLOv8-based adaptive head method; not named YOLOv8-D |
| DASSF-YOLOv8n | DASSF | YOLOv8n with dynamic upsampling and DyHead; not named YOLOv8-D |
| DyCAF-Net | DyCAF-Net | YOLOv8 with dynamic neck and class-aware head adaptation; not named YOLOv8-D |
| YOLOv8 + SGDM | SGDM-enhanced YOLOv8 | YOLOv8n with dynamic convolution plug-ins; not named YOLOv8-D |
A persistent misconception is that all “dynamic YOLOv8” models describe the same architectural idea. The literature instead splits into at least two distinct meanings. One meaning is continual-learning adaptation over time, which is the exact sense of YOLOv8-D in the VRU paper. The other meaning is input-conditioned feature processing inside the detector, as in dynamic gating, dynamic convolution, dynamic upsampling, or adaptive heads. Treating these as synonymous obscures substantial methodological differences.
2. Exact model definition in the literal YOLOv8-D paper
In the exact-naming paper, YOLOv8-D is introduced as a modified YOLOv8x-based continual-learning detection system for pedestrians, people standing by roadsides, bicyclists, and tricycles. The baseline detector is selected after comparison with Faster R-CNN, YOLOv5s, YOLOv5x, YOLOv7s, YOLOv7x, YOLOv8s, and YOLOv8x, after which YOLOv8x is chosen for its overall balance of accuracy and real-time performance. The paper attributes the underlying YOLOv8 baseline to background properties such as mosaic augmentation, anchor-free detection, the C2f module, a decoupled head, CIoU and DFL losses, and Adam for convergence, but these are not new contributions of YOLOv8-D itself (Nasir et al., 15 Jul 2025).
What turns YOLOv8x into YOLOv8-D is described narrowly. The paper states that it incorporates continual learning into YOLOv8 using gradient-based methods and, in practice, the only concrete mechanism documented is changing the optimizer from Adam to SGD during sequential adaptation to a new dataset. The model is first trained on VisDrone2019 and then adapted to Caltech Pedestrian. The authors interpret SGD’s inherent randomness as helping the model adapt to new data while retaining knowledge from previously learned tasks. The paper does not describe replay buffers, rehearsal memory, distillation loss, EWC/Fisher regularization, synaptic intelligence penalties, dynamic heads, parameter-freezing schedules, domain-adaptation blocks, auxiliary branches, modified YOLO loss terms, or custom optimizer update equations.
This definition is important because the model name suggests a richer dynamic architecture than the method section actually specifies. The paper presents YOLOv8-D as an adaptive system, but the documented technical novelty is optimizer-level continual-learning integration rather than a newly detailed dynamic backbone, neck, or head. A plausible implication is that the title’s “dynamic” should be read as dynamic adaptation across changing road environments rather than dynamic network topology.
The mathematical content in the paper is correspondingly limited. It provides evaluation equations for precision, recall, mean average precision, F1 score, and computational time:
and
No explicit continual-learning objective, no total loss decomposition for YOLOv8-D, and no SGD hyperparameterization are provided.
3. Data regime, sequential training protocol, and domain shift
The experimental narrative relies on two statistically different datasets. VisDrone2019 is used with 10,209 static images split into 6000 training, 2209 validation, and 2000 testing images, and is re-annotated to focus on VRU classes: pedestrian, people standing by roadsides (“people”), bicyclists or bicycle, and tricycles. Caltech Pedestrian is described as containing 350,000 bounding boxes annotated for about 2300 distinct pedestrians, with a split of 5000 training, 1500 validation, and 500 testing images in the reported experiments (Nasir et al., 15 Jul 2025).
The paper treats the datasets as statistically different in two explicit ways. First, they differ in image viewpoint: VisDrone is aerial or drone-view, whereas Caltech is street-view. Second, they differ in number of classes and class statistics: VisDrone, after re-annotation, contains multiple VRU categories with more tricycle and bicycle examples, whereas Caltech has many more pedestrian and people examples but far fewer bicycle and tricycle examples. This difference is central to the paper’s catastrophic-forgetting claim.
The validation or evaluation statistics reinforce that framing. For VisDrone, the paper reports 2209 images and 102,083 instances, comprising 50,600 pedestrians, 40,230 people, 10,933 tricycles, and 320 bicycles. For Caltech, it reports 1500 images and 159,700 instances, comprising 85,000 pedestrians, 72,000 people, 1,500 tricycles, and 1,200 bicycles. The class imbalance across domains is thus not incidental; it is the empirical substrate on which the continual-learning argument is built.
The training setup is only partially specified. The main benchmark experiments are reported at 300 epochs, with a confidence threshold of 0.20 and mAP@0.50 as the principal detection metric. The paper does not provide image resolution, batch size, optimizer hyperparameters, augmentation settings beyond general YOLOv8 background mention, hardware specifications, inference batch size, or NMS threshold. This under-specification is methodologically significant because the claimed continual-learning effect is tied to optimizer choice, yet the optimizer settings themselves are not reported.
4. Quantitative results and real-time profile
The initial detector-selection experiment identifies YOLOv8x as the strongest baseline among the evaluated detectors on the VisDrone-based VRU task (Nasir et al., 15 Jul 2025).
| Model | Precision / Recall / [email protected] / F1 | FPS / Inference time |
|---|---|---|
| Faster R-CNN | 0.55 / 0.67 / 0.523 / 0.601 | 4.55 / 220 ms |
| YOLOv5s | 0.455 / 0.292 / 0.291 / 0.343 | 175 / 5.7 ms |
| YOLOv5x | 0.582 / 0.381 / 0.353 / 0.412 | 24 / 41.1 ms |
| YOLOv7s | 0.545 / 0.342 / 0.219 / 0.341 | 290 / 12.6 ms |
| YOLOv7x | 0.571 / 0.215 / 0.225 / 0.381 | 46 / 21.5 ms |
| YOLOv8s | 0.568 / 0.335 / 0.357 / 0.414 | 625 / 1.6 ms |
| YOLOv8x | 0.763 / 0.485 / 0.514 / 0.462 | 101 / 9.9 ms |
The abstract reports that, relative to YOLOv5x, YOLOv8x improves F1 by 12.14% and mAP by 45.61%; relative to YOLOv7x, the improvements are 21.26% in F1 and 128.44% in mAP. Those percentages are numerically consistent with the reported table values.
The central evidence for YOLOv8-D then comes from sequential VisDrone-to-Caltech adaptation. The paper compares four states: YOLOv8x on VisDrone, YOLOv8x trained from scratch on Caltech, YOLOv8x adapted from VisDrone to Caltech with Adam, and YOLOv8x adapted from VisDrone to Caltech with SGD.
| Setting | Precision / Recall / [email protected] / F1 | Training time |
|---|---|---|
| YOLOv8x(a) VisDrone | 0.763 / 0.485 / 0.5415 / 0.462 | 15.831 h |
| YOLOv8x(b) Caltech from scratch | 0.852 / 0.776 / 0.475 / 0.412 | 10.72 h |
| YOLOv8x(c) VisDrone→Caltech with Adam | 0.855 / 0.772 / 0.461 / 0.441 | 10.53 h |
| YOLOv8x(d) VisDrone→Caltech with SGD | 0.889 / 0.783 / 0.608 / 0.534 | 16.98 h |
The authors interpret the similarity between Caltech-from-scratch training and VisDrone→Caltech adaptation with Adam as evidence of catastrophic forgetting. By contrast, the SGD-based sequential version improves precision from 0.855 to 0.889, recall from 0.772 to 0.783, mAP from 0.461 to 0.608, and F1 from 0.441 to 0.534 relative to the Adam sequential baseline. The abstract’s “21.08% improvement in F1 score and 31.86% improvement in mAP” corresponds most faithfully to that Adam-versus-SGD sequential comparison, because and .
The real-time claim rests on the detector-selection benchmark rather than on a special YOLOv8-D inference study. For YOLOv8x, the paper reports 101 FPS and 9.9 ms inference time per image. Using the paper’s own definition,
one second of 30 FPS input corresponds to 0.297 s of computation for YOLOv8x. The principal trade-off introduced by YOLOv8-D is therefore longer training rather than slower inference: the SGD continual-learning variant increases reported training time from 10.53 h to 16.98 h relative to the Adam sequential baseline.
5. Broader dynamic YOLOv8 family
Outside the exact-naming paper, the phrase “YOLOv8-Dynamic” is often used informally for architectures that make YOLOv8 input-adaptive inside the detector itself. The most directly comparable example is YOLO-DS, a YOLOv8-based detector that preserves the standard YOLOv8 scaling rules across N, S, M, L, and X while inserting Dual-Statistic Synergy Gating (DSG) and Multi-Path Segmented Gating (MSG) into the C2f structure. Its Dual-Statistic Synergy Operator models each channel with the channel-wise mean and the peak-to-mean difference, and YOLO-DS reports AP gains of 1.1% to 1.7% over YOLOv8 across the five canonical scales on MS-COCO with only minimal latency increases (Huang et al., 26 Jan 2026).
A second line of work uses dynamic or adaptive components without standardizing the name. The traffic-sign recognition study proposes an enhanced YOLOv8 with Coordinate Attention, BiFPN, a P2 160×160 small-object detection layer, BOTNet-style MHSA, ODConv, LSKA, and losses involving EIoU, WIoU, Focal Loss, and BCE-With-Logits. It reports 91.5% mAP, improvement over YOLOv5 at 87.3%, and 45 FPS, but it explicitly does not define the model as YOLOv8-Dynamic or YOLOv8-D (Ibrahim et al., 2 Apr 2025).
A third line centers the dynamic component in the head. ADA-YOLO replaces the standard YOLOv8 detection head with an Adaptive Head containing Dynamic Visual Feature Localisation (DVF), multi-scale convolution layers, and a Joint-Guided Regression Module (JGRM). On BCCD, it reports mAP@50 = 0.912 versus 0.888 for YOLOv8, mAP@50-95 = 0.630 versus 0.602, and a model-size reduction from 26.9 MB to 8.7 MB, but it is not named YOLOv8-D (Liu et al., 2023).
A fourth line modifies neck and head jointly for small-object aerial detection. DASSF-YOLOv8n introduces DSSFF with dynamic upsampling via DySample, adds a 160×160 x-small detection head, and applies DyHead with scale-aware, spatial-aware, and task-aware attention. Relative to YOLOv8n, it reports +9.2 mAP50 on VisDrone-2019 and +2.4 mAP50 on DIOR (Li et al., 2024).
A fifth line makes the neck itself dynamic. DyCAF-Net replaces YOLOv8’s PANet neck with an input-conditioned equilibrium-based fusion neck, adds dual dynamic attention, and augments the detection heads with class-aware feature adaptation. It maintains roughly 11.1M parameters and improves YOLOv8 on 10/13 datasets in mAP@50-95 according to the paper’s summary, but again the official model name is DyCAF-Net rather than YOLOv8-D (Jahin et al., 5 Aug 2025).
A sixth line uses plug-in dynamic convolution. SGDM inserts a Static-Guided Dynamic Module before the P3, P4, and P5 outputs feeding the detection head in YOLOv8n. On COCO val2017, the reported change is AP 36.4 → 38.1, Params 3.1M → 3.3M, FLOPs 8.7G → 9.1G, and Time 4.2 ms → 4.3 ms (Xing et al., 2024).
Taken together, these papers show that “dynamic YOLOv8” is not a single architecture class. It spans at least four mechanisms: continual-learning adaptation across datasets, sample-dependent gating inside C2f blocks, dynamic or adaptive heads, and dynamic neck fusion or convolutional plug-ins. This suggests that the phrase is best treated as a descriptive umbrella rather than a stable canonical model family.
6. Misconceptions, limitations, and interpretive boundaries
The first misconception is nominal. YOLOv8-Dynamic (YOLOv8-D) is not a generally accepted label for every adaptive YOLOv8 variant. In the literature summarized here, only one paper uses that exact model name; the others use distinct names such as YOLO-DS, ADA-YOLO, DASSF, DyCAF-Net, or SGDM-enhanced YOLOv8 (Nasir et al., 15 Jul 2025, Huang et al., 26 Jan 2026, Liu et al., 2023, Li et al., 2024, Jahin et al., 5 Aug 2025, Xing et al., 2024).
The second misconception is architectural. In the exact YOLOv8-D paper, “dynamic” does not mean a new dynamic head, dynamic kernel-selection mechanism, or explicit input-conditioned routing. The documented mechanism is optimizer-level continual-learning integration, specifically switching from Adam to SGD during sequential adaptation. A plausible implication is that the title overstates architectural novelty relative to the method description.
The third misconception concerns evidential scope. The exact YOLOv8-D paper demonstrates sequential adaptation only across two datasets and one adaptation direction, with no replay buffer, no explicit continual-learning regularizer, no distillation term, and no custom continual-learning objective. The evidence is therefore empirical and comparatively narrow. Similarly, several related dynamic-YOLOv8 papers are strong concept demonstrations but incompletely specified for exact reproduction: the traffic-sign model does not provide an exact layer-by-layer YAML or clean total loss formula; ADA-YOLO leaves the exact YOLOv8 variant and several head details unspecified; DASSF omits some training hyperparameters; DyCAF-Net under-specifies equilibrium-solver details; and SGDM contains minor equation ambiguities in its branch concatenation formula (Ibrahim et al., 2 Apr 2025, Liu et al., 2023, Li et al., 2024, Jahin et al., 5 Aug 2025, Xing et al., 2024).
The most defensible synthesis is therefore narrow and two-part. In the literal sense, YOLOv8-Dynamic is the VRU-oriented continual-learning extension of YOLOv8x introduced for VisDrone-to-Caltech adaptation. In the functional sense, the phrase can also denote a broader research direction in which YOLOv8 is made adaptive through dynamic gating, dynamic convolution, adaptive heads, or dynamic fusion. Maintaining that distinction is necessary for accurate citation, replication, and comparison across the YOLOv8 literature.
