Illustrative convolution and pooling feature stacks; not the exact trained architecture, dimensions or model output. — CNN feature-stack illustration · not model output

UT AUSTIN · DEEP LEARNINGacademic

One image. Two prediction tasks.

I built a multi-task encoder–decoder CNN with skip connections to produce pixel-level segmentation and depth regression from a single image.

Shared image input · segmentation + depth

What I worked on

The model performs both a classification task at the pixel level and a continuous depth-regression task. Skip connections are part of the encoder–decoder architecture.

Severe class imbalance required an explicit change to the objective: I used class weighting rather than treating all classes equally.

Tools and methods PyTorch · Encoder–decoder CNN · Skip connections · Class-weighted objective