PyTorch-AI-road-detection

Artificial intelligence model, training loop, concurrent model loader, inference
From now image processing runs concurrently
Full test with car detection https://www.youtube.com/watch?v=VmhW1dGFPMw&t=1s&ab_channel=SamuelBachorik

Custom AI model trained on custom dataset (1000 photos).
GPU used = RTX 3060 12GB
Example:

Model

AI Model created with PyTorch.
The model is being trained on GPU, model's device is set to "CUDA"

Model architecture

• Encoder - decoder model
• 16x Conv2d layers divided in encoders and decoder
• ReLu activatios + BatchNorm2d
• All of this in nn.Sequential
• Kernel 3x3
• 17 360 898 Model Parameters

The course of the loss function

90-110 epoch is enough for this model with this dataset.
For this project I used mean squared error loss function = ((y - y_pred) ** 2).mean()

Dataset

Dataset consists of 1300 photos from rainy and sunny city. 50% rainy / 50% sunny photos
Link to dataset photos - Dataset
Example:

How to use

In Run_training.py set folders with downloaded dataset like this

folders_training.append("C:/Users/Samuel/PycharmProjects/Conda/City_dataset/City_sunny1/")
folders_training.append("C:/Users/Samuel/PycharmProjects/Conda/City_dataset/City_sunny2/")
folders_training.append("C:/Users/Samuel/PycharmProjects/Conda/City_dataset/City_rainy/")
folders_training.append("C:/Users/Samuel/PycharmProjects/Conda/City_dataset/City_rainy2/")
folders_training.append("C:/Users/Samuel/PycharmProjects/Conda/City_dataset/City_2/")

Then Run Run_training.py

When model is trained you can run inference like this -
In segmentation_inference set saved model Path

# Path to trained weights
self.PATH = "./Model1"

In Run_video_inference.py set your desired video

cap = cv2.VideoCapture("C:/Users/Samuel/PycharmProjects/Condapytorch/City.mp4")

Run Run_video_inference.py
When inference is done you will get output.avi video.

Run with pretrained model

If you want to run inference with pretrained model you need to download model trained wieghts -
You can choose - Models weights

IMPORTANT: Make sure your downloaded model weights corresponds with Model architecture.
Model_weights_1 for Model 1
Model_weights_2 for Model 2

In segmentation_inference set model Path

# Path to trained weights
self.PATH = "Model_weights_1.pth"

In Run_video_inference.py set your desired video

cap = cv2.VideoCapture("C:/Users/Samuel/PycharmProjects/Condapytorch/City.mp4")

Run Run_video_inference.py
When inference is done you will get output.avi video.

Test video

For best result use video with aspect ratio 4:3 and ressolution 1024x768
You can test model on your own video or you can download here - Download Video