YOLOv8 object detection example

• Demo
• Dataset
• AP
• Classes
• Training Log
• Weights
• Environment
• Tuning_Hyperparameters
• Speed
• Google Colab
• References

Demo

Dataset

https://drive.google.com/file/d/1rlDFnYgzLtDNS_tqJDE0-LfURDUtXwsn/view?usp=sharing

AP

             Class     Images  Instances      Box(P          R      mAP50  mAP50-95)
               all       2091      15400      0.726      0.584      0.643      0.367
               car       2091       9024      0.804      0.777       0.83      0.536
             truck       2091        873      0.805      0.652      0.737      0.493
        pedestrian       2091       1638       0.65      0.516      0.572      0.275
         bicyclist       2091        151      0.665      0.503      0.556      0.303
             light       2091       1747      0.833      0.682      0.739      0.377
           pothole       2091       1967      0.598      0.376      0.424      0.216

Classes

1. car: with 68,008 labels
2. truck: with 4,170 labels
3. pedestrian: with 8,590 labels
4. bicyclist: with 955 labels
5. light: with 8,723 labels
6. pothole: with 10,024 labels

• total: 100,470 labels
• Train : valid = 9:1

Training Log

|

Weights

trained with 300 epochs
https://drive.google.com/file/d/1-mfFLoZrUd6jlrmd9V9ym5_g2J3SIyPV/view?usp=sharing

Environment

• VM: Google Colaboratory
• GPU: NVIDIA T4 Tensor GPU
• NVIDIA-SMI 525.105.17 Driver Version: 525.105.17 CUDA Version: 12.0
• nvcc: NVIDIA (R) Cuda compiler driver
• Cuda compilation tools, release 11.8, V11.8.89
• Build cuda_11.8.r11.8/compiler.31833905_0

Tuning Hyperparameters

https://colab.research.google.com/drive/1DlWlXhY4k5pwOVGWSPEt6zYDLj4sLH39#scrollTo=LlPmSN513UYT

Speed

network I/O : time of sending image from front page to inferencing server
inferencing time : time of detection of objects

• above chart represent results of processing 10 images with models trained only for 50 epochs

Google Colab

Link

References

• Training data1(self-driving) : https://www.kaggle.com/alincijov/self-driving-cars
• Training data2(pothole) : https://www.dropbox.com/s/qvglw8pqo16769f/pothole_dataset_v8.zip?dl=1
• The dataset used for training was preprocessed with above 2 datasets(reducing size, cross labeling, balancing number of labels of each class)
• test videos
  https://drive.google.com/file/d/1uYKr_ogmZp1U8akz3qWKsCov6hu9Voob/view?usp=sharing
  https://drive.google.com/file/d/1IEPLROgmtc7MTDinsMRcIyaeaweNiBzU/view?usp=sharing