Assignments and project from the course TDT4265 Computer vision and Deep Learning @ NTNU.
Simple single layer neural network used classifying handwritten digits on the MNIST dataset.
More complex multilayer neural network for classifying handwritten digits on the MNIST dataset.
Convolutional neural networks (CNNs) and pretrained CNNs used for classification of various objects in images.
Object detection using a single shot multibox detector (SSD).
Detecting wear and tear on roads using a single shot detector. Project related to the course TDT4265 - Computer Vision and Deep Learning @ NTNU.
This Project folder contains code for a single shot multibox detector trained to detect different types of road damages on Norwegian roads. The folder contains two SSD models
-
A model that has been pretrained on the RDD2020 dataset, a dataset containing 26000 labeled images from around the world. The pretraining has been done for 16k iterations lasting 8 hours on a NVIDIA T4 card. The pretrained model has been further trained on a private dataset containing labeled images from Norwegian roads. This model has been trained for another 18k iterations, lasting 11 hours on a NVIDIA T4 card. The final model reached a mAP of 0.03, comfortably beating the baseline target model (0.028).
-
A Feature-Fused SSD model, which has been developed to further increase performance on the RDD2020 dataset. The model reaced a mAP of 0.44 after 20k iterations, significantly exceeding the baseline target model (0.36 mAP)
Install the required dependencies
pip install -r requirements.txt
Navigate to the SSD folder
cd BaseModel/SSD
To train the model, simply execute
python train.py configs/<config_name>
where <config_name> is replaced by one of the following:
train_tdt4265.yamlfor training the first model on the Norwegian dataset (given you have access).train_rdd2020_server.yamlfor training the first model on the RDD2020 dataset.train_rdd2020_fused_ssd_server.yamlfor training the Fused SSD on the RDD2020 dataset.