Systems · 2025

Deep-Learning Vision

A set of computer-vision experiments that took me from training CNNs from scratch to working with large pretrained vision and vision-language models.

PyTorch · CUDA · CNN · segmentation

What I explored

I trained a custom CNN on CIFAR-10 and built a VGG16 encoder-decoder for semantic segmentation on Cityscapes, including data loading, augmentation, GPU training, inference, and evaluation with IoU and confusion matrices.

I also explored DINOv2 attention maps and used CLIP for zero-shot classification.

Outcome

CLIP reached 92.72% zero-shot accuracy across 10,000 CIFAR images. More importantly, the project let me compare how traditional supervised pipelines and large pretrained representations behave on the same kinds of visual tasks.