robertklee / kitti-roadseg

A course project for road segmentation using a U-Net Convolutional Neural Network on the KITTI ROAD 2013 dataset

Welcome to my GitHub!

👋 I am an award-winning and ambitious B.Eng. in electrical/computer engineering passionate about:

• machine learning (specifically computer vision problems)
• software engineering
• firmware engineering and embedded systems

💼 Workplace: Microsoft Software Engineer II.

I currently work on vector search within Azure Search team.

👨‍💻 I'm currently working on:

• building and hosting my own personal website (check it out here!)
• creating a computer vision project for 3D bounding box for autonomous vehicle applications

✅ Some of my most exciting projects include:

• Human Pose Estimation on COCO 2017 Dataset: I led a team to design and train a deep neural network for human pose estimation (labelling various joint locations). Using a modified stacked hourglass network and a newly implemented data ingestion and preprocessing pipeline, we achieved results comparable with state-of-the-art from late 2017, in less than three months of work.

• Road Segmentation Machine Learning Project on KITTI ROAD dataset: A convolutional neural network (CNN) for semantic segmentation of road surfaces within a driving context.

• Embedded C Optimization Project with Discrete Cosine Transform algorithm: Implementation of DCT algorithm in C and optimizing performance using more efficient software algorithms, software-level optimizations, and hypothetical firmware- and hardware-based optimizations.

• Computer Vision Project on Monocular Depth Estimation: A CNN for depth estimation on DrivingStereo dataset using Keras & Python.

📝 Reach out to me on LinkedIn if you want to collaborate or connect!

kitti-roadseg's Issues

Query

How to predict on the road image data using your model from model.py using this weights "weights_epoch20_val_loss-1.9479_train_loss_-1.9623.hdf5" available in your one drive link?

Also, are these weights fully trained on the KITTI dataset?

some queries

hello author,

great work with the keras tensorflow implementation. i have some doubts regarding your implementation.

1. why is the DEFAULT_EPOCHS = 20 and DEFAULT_BATCH_SIZE = 12. is it because of small dataset?
2. image size = 640x192, did you match this with the image generator function?
3. how can i implement a simple encoder decoder using vgg 16.

thank you

tensorflow gpu

hello again,

did you use tensorflow without gpu? because your requirements.txt uses cpu version of tensorflow right?

thank you