EVALUATING YOLOV8 EFFICIENCY ON GPU, CPU AND RASPBERRY PI: STANDARD VS. SEPARABLE CONVOLUTIONS

> Author: Pablo Vicente Reyes Pino

Code and method Full README

This page shows the project's results. The methodology, the design decisions and the limitations are documented in the repository's README.

1.2 Problem statement

Figure 1: Object detection example (YOLO) — <p align="center"><em>Figure 1: Object detection example (YOLO). Places object detection within the family of computer-vision tasks: semantic segmentation, classification + localization, object detection and instance segmentation.</em></p>

3.2 YOLOv8 architecture

Figure 2: YOLOv8 architecture (Backbone-Neck-Head) — <p align="center"><em>Figure 2: Standard YOLOv8 model as implemented (Terven et al., 2023). 640×640×3 input, four Backbone stages, multi-scale fusion in the Neck and three decoupled prediction branches in the Head.</em></p>

3.3 Conv2D vs. separable convolutions

Figure 3: Conv2D vs SeparableConv2D — <p align="center"><em>Figure 3: Conv2D vs SeparableConv2D (comparative diagram).</em></p>

Figure 4: Standard vs separable convolution processes — <p align="center"><em>Figure 4: Example of standard vs. separable convolution processes. Above, the single 3D filter sweeping space and channels simultaneously; below, the decomposition into per-channel filtering (depthwise) followed by cross-channel combination (pointwise).</em></p>

4.1 Research design

Figure 5: Training/validation process — <p align="center"><em>Figure 5: Training / validation process.</em></p>

5.3 Performance evolution: mAP every 10 epochs

Figure 6: mAP evolution every 10 epochs — <p align="center"><em>Figure 6: Evolution of the mAP metric every 10 epochs. (a) YOLOv8-Conv2D model. (b) YOLOv8-Separable model.</em></p>

5.4 Validation results per model

Figure 7: YOLOv8-Conv2D detection model trained for 50 epochs — <p align="center"><em>Figure 7: Detection with YOLOv8-Conv2D trained for 50 epochs. Ground-truth boxes in red, model predictions in yellow, over COCO2017 validation images.</em></p>

Figure 8: YOLOv8-Separable detection model trained for 100 epochs — <p align="center"><em>Figure 8: Detection with YOLOv8-Separable trained for 100 epochs. Same visualization protocol and same validation images as Figure 7, for direct comparison.</em></p>

5.7 Visualizations derived from the results

Training convergence — YOLOv8-Separable — This curve is not a summary statistic: it is the epoch-by-epoch box_loss / class_loss / val_loss sequence saved in the output cell of 02_yolov8_separable_conv.ipynb itself (Epoch 1/10 … Epoch 10/10), included so the reported metrics can be audited against the raw training log and not only against the final numbers.