YOLO advances to its genesis: a decadal and comprehensive review of the You Only Look Once (YOLO) series

In plain words

YOLO is a family of object detectors. It looks at a whole image once and, in a single pass, says where each object is and what it is. That single pass is why it is fast enough for video, cameras on vehicles and factory lines.

This review follows the family from its first version to the most recent one, going backwards in time. For each version it asks what changed and why: faster, more accurate or lighter. It also lists where each version still struggles, such as tiny objects, crowded scenes and fixed grids. The authors then survey how YOLO is used in five areas: driving and traffic, medical imaging, factories, security and farming. The last part looks ahead to running YOLO on small edge devices and combining it with language models.

The review does not train or test any model of its own. Its value is the map of the family and the caveats for comparing its versions.

TL;DR

• Review article (no new method, no experiments) covering YOLOv1 to YOLOv12 and five variant detectors [abstract].
• Each version is described by its architectural changes, reviewed in reverse chronological order [abstract] fig. 2.
• Gives per-version limitations, five application areas, and future directions in edge deployment, embodied AI and LLM integration §7 §8.
• Caveat: the review itself reports no benchmark; every number is taken from other papers.

Problem

Object detection must find and label objects in an image at a speed the application can accept p. 2. Before deep learning, detectors used hand-crafted features and classifiers, such as the Viola–Jones detector and HOG p. 3. Two-stage detectors such as R-CNN and Faster R-CNN separate proposal generation from classification, which is slow for real-time use p. 5–6. YOLO was introduced to merge region proposal and classification into one network §1.4 p. 6.

The review states the challenges it wants to address: real-world variability, illumination, occlusion and clutter, and speed and efficiency p. 2.

Method

The review is organised as follows §1.5 fig. 2:

1. Background on object detection and the traditional methods §1.1 §1.2.
2. The timeline of detector paradigms and the YOLO trajectory from v1 to v12 §1.3 §1.4 fig. 1 §2.
3. Context and distinctions from prior YOLO literature §3.
4. Review of each YOLO version and its variants, reverse chronologically [abstract] §4.
5. Applications in five areas [abstract] §5.
6. Challenges, limitations and future directions §6.
7. Future directions: edge and IoT deployment §7.1, embodied AI §7.2.
8. Expanding YOLO into broader AI: AGI §8.1 and LLM integration §8.2.
9. Environmental impact §9 and conclusion §10.

The abstract describes the procedure as a reverse chronological analysis of each version's contributions to speed, accuracy and computational efficiency [abstract]. The card covers §1, §6–§10 and the abstract. §2–§5, which hold the per-version reviews, were not summarised here (see Pitfalls).

Version details stated in the read pages §1.4 p. 6:

• YOLOv1: a single network that predicts bounding boxes and class probabilities directly for each grid cell §1.4.
• YOLOv2: DarkNet-19 backbone; anchor boxes refined by K-means clustering §1.4.
• YOLOv3: DarkNet-53 backbone §1.4.
• YOLOv4: CSPDarkNet-53 backbone, PANet neck, mosaic data augmentation §1.4.
• YOLOv10: lightweight classification head and distinct spatial and channel transformations §1.4.
• YOLOv11: C3k2 block in the backbone and C2PSA for spatial attention §1.4.
• YOLOv12: attention-centric design with an Area Attention (A²) module, residual efficient layer aggregation networks (R-ELAN), FlashAttention and adjusted MLP ratios §1.4.

Figure 1 is a timeline of detector paradigms, placing YOLO versions against two-stage, anchor-free and transformer detectors fig. 1. Figure 2 is a mind map of the manuscript fig. 2.

No algorithm or equations are presented: this is a review article. The one procedure the card sketches is the single-pass idea that the review attributes to YOLOv1 §1.4, with the two-boxes-per-cell limit stated later p. 63–64:

# YOLOv1-style single pass  [§1.4] [p. 63–64]
# Input:  image (H, W, 3)
# Output: per grid cell, boxes with class probabilities

def detect(image):
    # Run one CNN over the whole image, no separate proposal step  [§1.4]
    features = backbone(image)              # (S, S, C) grid of cells
    # Each cell predicts a fixed number of boxes and class scores  [p. 63–64]
    predictions = head(features)            # (S, S, boxes * 5 + classes)
    return predictions

Hyperparameters

Not applicable. The review states no training or inference hyperparameters in the sections read §1.5. Per-version values are in the cited papers, not in this card.

Results

Author-reported review claims. The review runs no experiments of its own. Data availability: "No datasets were generated or analysed" p. 68.

Paper vs code

No official code. The abstract, the read pages and the arXiv metadata name no repository for this review. The repositories the review cites for individual YOLO versions (e.g. Ultralytics documentation p. 72) are third-party or vendor sources, not code for this paper.

Pitfalls

• Coverage gap in this card. The per-version reviews (§2–§5) were not read in this compilation. Anyone implementing a specific version should read the original paper for that version, not this review.
• Version list mismatch. The keywords name YOLOv13 to YOLOv20 [abstract], but the abstract and body stop at YOLOv12. Treat the keywords as unreviewed.
• No benchmark, no ranking. The review gives no table that compares versions on the same data. Its statements on speed or accuracy are carried over from the cited papers, which use different datasets and hardware.
• Self-citation. Several cited application papers share authors with this review, including the first and last author p. 80–81. Check independent sources before relying on a cited claim.
• Reference mix. The reference list includes peer-reviewed venues, preprints and vendor documentation p. 70–72 p. 80–81. Treat them with different weight.
• Sensitive trade-offs stated in the review. YOLOv4 has a high computational demand that restricts deployment on low-power devices p. 63. YOLOv1's fixed structure of two boxes per grid cell makes small or closely grouped objects hard to detect p. 63–64.
• Metrics caveat. The review notes that a single metric does not capture speed, accuracy and computational cost together, and recommends reporting several p. 64–65.
• Environmental cost. The review states that training and retraining YOLO is energy-intensive and cites energy, water and carbon effects §9.

Lineage

• Builds on: traditional detectors (Viola–Jones, HOG, DPM) and two-stage detectors (R-CNN, Fast R-CNN, Faster R-CNN) §1.1 §1.3 p. 3–6. No card for these papers exists.
• Compared against: the non-YOLO detectors on the timeline in fig. 1.
• Superseded by: not applicable. The review itself names later versions beyond v12 only in its keywords.