CoChat

T-YOLO: Tiny Vehicle Detection Based on YOLO and Multi-Scale Convolutional Neural Networks

Open access
Padilla Carrasco, Daniel, Rashwan, Hatem A., García, Miguel Ángel, Puig, Domènec · 2023 · IEEE Access

The DOI matches its Crossref record, with a small difference in title or year; 40 of the 40 references it cites were found in a registry or an index; 3 of the 11 sources its claims cite were checked for support.

Abstract

To solve real-life problems for different smart city applications, using deep Neural Network, such as parking occupancy detection, requires fine-tuning of these networks. For large parking, it is desirable to use a cenital-plane camera located at a high distance that allows the monitoring of the entire parking space or a large parking area with only one camera. Today’s most popular object detection models, such as YOLO, achieve good precision scores at real-time speed. However, if we use our own data different from that of the general-purpose datasets, such as COCO and ImageNet, we have a large margin for improvisation. In this paper, we propose a modified, yet lightweight, deep object detection model based on the YOLO-v5 architecture. The proposed model can detect large, small, and tiny objects. Specifically, we propose the use of a multi-scale mechanism to learn deep discriminative feature representations at different scales and automatically determine the most suitable scales for detecting objects in a scene (i.e., in our case vehicles). The proposed multi-scale module reduces the number of trainable parameters compared to the original YOLO-v5 architecture. The experimental results also demonstrate that precision is improved by a large margin. In fact, as shown in the experiments, the results show a small reduction from 7.28 million parameters of the YOLO-v5-S profile to 7.26 million parameters in our model. In addition, we reduced the detection speed by inferring 30 fps compared to the YOLO-v5-L/X profiles. In addition, the tiny vehicle detection performance was significantly improved by 33% compared to the YOLO-v5-X profile.

Overview

CoChat Summary

## TL;DR

The paper proposes T-YOLO, a lightweight modification of YOLOv5 for tiny vehicle detection in cenital-view parking images, using a multi-scale module and attention mechanisms. On PKLot, the adapted model improved precision substantially over the baseline, reaching up to 96.34% precision versus 63.87% for baseline YOLOv5, while keeping roughly 30 fps.

## Key findings

  • Replacing YOLOv5’s Focus layer with the proposed multi-scale module (MSM) improved precision from 0.6273 in YOLOv5s to 0.9203 in YOLOv5s_MSM.
  • The best reported model, Yolov5s_MSM_CAM, achieved a precision of up to 96.34%; the paper reports the baseline YOLOv5 precision as 63.87%.
  • The paper reports a small reduction in trainable parameters, from 7.28 million in YOLOv5-S to 7.26 million in the proposed model.
  • The proposed adaptation ran at about 30 fps; the single-branch solution Yolov5_SM_x2 increased speed from 35 fps to 42 fps and was described as outperforming YOLOv5m in speed.

## Methods

  • Design: The study modifies YOLOv5 by replacing the Focus layer with a multi-scale module based on ENet-style branches (x1, x2, x4) and by adding spatial/channel attention modules (SAM, CAM, SCAM) in different configurations.
  • Data/sample: PKLot was used; the authors selected the cenital-plane subset (PUCPR) with 4,474 images and about 100 parking spots per image, totaling 424,269 tagged spots. The data were split 80% train and 20% test, and annotations were translated to COCO format.
  • Analysis: The authors conducted ablation experiments comparing baseline YOLOv5, fine-tuned YOLOv5, MSM variants, attention variants, and pooling variants, evaluating precision, recall, mAP, inference speed (fps), and parameter count.

## Limitations

  • The authors state that the model is designed for a specific application: tiny cars from a cenital-view camera in large parking areas, so generalization beyond this scenario is not demonstrated in the text.
  • The paper notes that using a multi-scale approach can add redundancy and may degrade performance if not chosen carefully; they found the three-branch MSM may not be the most efficient for this dataset.
  • Evident from the text: the dataset is restricted to a monitored/tagged region of PKLot, and the authors apply a mask to exclude nontagged cars, which limits what the model is evaluated on.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/access.2021.3137638

Cites

40 of 40 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

FIGURE 1 Overview of desired large parking configuration and a cential view camera.
FIGURE 1 Overview of desired large parking configuration and a cential view camera.CC-BY · From the PDF
FIGURE 2 ENet initial block architecture.
FIGURE 2 ENet initial block architecture.CC-BY · From the PDF
FIGURE 3 Multi-scale module architecture.
FIGURE 3 Multi-scale module architecture.CC-BY · From the PDF
FIGURE 4 Standard focus layer of YOLO-v5 model.
FIGURE 4 Standard focus layer of YOLO-v5 model.CC-BY · From the PDF
FIGURE 5 Spatial and channel attention modules.
FIGURE 5 Spatial and channel attention modules.CC-BY · From the PDF
FIGURE 6 Proposed architectures based on Yolo-v5. a) Yolo-v5 baseline architecture. b) Yolo-v5 with replacing focus layer by multi-scale module (MSM). c) Yolo-v5 with MSM and spatial and channel attention module (SCAM) at the end of backbone. d) Merging SCAM into each branch from the MSM layer.
FIGURE 6 Proposed architectures based on Yolo-v5. a) Yolo-v5 baseline architecture. b) Yolo-v5 with replacing focus layer by multi-scale module (MSM). c) Yolo-v5 with MSM and spatial and channel attention module (SCAM) at the end of backbone. d) Merging SCAM into each branch from the MSM layer.CC-BY · From the PDF
FIGURE 7 An example of the PKLot dataset with the pretrianed YOLO-v5 model.
FIGURE 7 An example of the PKLot dataset with the pretrianed YOLO-v5 model.CC-BY · From the PDF
FIGURE 8 An example of the PKLot dataset with the fine-tuned YOLO-v5 model.
FIGURE 8 An example of the PKLot dataset with the fine-tuned YOLO-v5 model.CC-BY · From the PDF
FIGURE 9 An example of the PKLot dataset with the YOLO-v5 model with multi-scale module (MSM).
FIGURE 9 An example of the PKLot dataset with the YOLO-v5 model with multi-scale module (MSM).CC-BY · From the PDF
FIGURE 10 Two examples of the PKLot dataset with (Row 1) original images, (Row 2) car detection results with the fine-tuned YOLO-v5 model with the PKLot dataset, (Row 4) car detection results with the Yolov 5 s _ MSM model, and (Row 5) car detection results with the Yolov 5 s _ MSM _ CAM _ backbone model. Green bounding-boxes are related to a detection with IoU > 0 : 95, and the red ones with otherwise.
FIGURE 10 Two examples of the PKLot dataset with (Row 1) original images, (Row 2) car detection results with the fine-tuned YOLO-v5 model with the PKLot dataset, (Row 4) car detection results with the Yolov 5 s _ MSM model, and (Row 5) car detection results with the Yolov 5 s _ MSM _ CAM _ backbone model. Green bounding-boxes are related to a detection with IoU > 0 : 95, and the red ones with otherwise.CC-BY · From the PDF
Image 11
CC-BY · From the PDF
Image 12
CC-BY · From the PDF
Image 13
CC-BY · From the PDF
Image 14
CC-BY · From the PDF
Image 15
CC-BY · From the PDF
Image 16
CC-BY · From the PDF
Image 17
CC-BY · From the PDF

Verification

Claims the paper makes

2 Valid · 5 Not valid · 7 Not checked · 53 Own results and interpretations

34 more claims

Create a free account to see all 67 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 22
  • Probably not a substantial claim. 9

Create a free account to see these claims.

Citations

Cites

Show all 40 references

Cited by

Metadata

Authors
Padilla Carrasco, D., Rashwan, H. A., García, M. Á., & Puig, D.
Affiliations
  • Department of Computer Engineering and Mathematics, Universitat Rovira i Virgili, Tarragona, Spain
  • Universitat Rovira i Virgili (ES)
  • Universidad Autónoma de Madrid (ES)
Venue
IEEE Access
Published
2023
Volume
11
Pages
22430-22440
ISSN
2169-3536 (ISSN-L)
OpenAlex
W4205762292
Cited by
145
References
37
Citations by year
2026: 22 · 2025: 40 · 2024: 54 · 2023: 25
FWCI
6.36
Type
journal article
Language
en
Open access
Locations
  • IEEE Access (publishedVersion, cc-by)
  • Biblos-e Archivo (Universidad Autónoma de Madrid) (submittedVersion, public-domain)
  • Rovira & Virgili repository (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Hispana (submittedVersion)
Funders
Agncia de Gesti dAjuts Universitaris i de Recerca (2018 DI 054)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Video Surveillance and Tracking Methods
Keywords
small object detection, vehicle detection, YOLOv5, multi-scale convolutional neural network, lightweight object detection model, parking occupancy detection, smart cities, parameter reduction, real-time inference

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.