Computer vision

DRIVABLE SPACE SEGMENTATION

A custom U-Net for drivable-space segmentation.

Problem

Drivable-space segmentation identifies the regions of an image that a vehicle may traverse.

Dataset & architecture

A custom U-Net trained for drivable-space segmentation using nuScenes and PyTorch.

Segmentation compare

Segmentation comparePlaceholder scene

Reveal the mask

Real input frames, ground-truth masks and model predictions replace this scene once supplied.

The existing record reports a custom U-Net using PyTorch and nuScenes, with 7.7M parameters, mIoU 0.9489 and 43+ FPS. These are reported figures, not independently reproduced results, and the measurement conditions are not yet published — they should not be used for comparisons.

Reported results

Parameters: 7.7M; Reported in the previous portfolio record; not independently reproduced. Evaluation conditions (dataset split, input resolution, IoU aggregation, hardware, precision, timing scope) are not yet documented.
mIoU: 0.9489; Reported in the previous portfolio record; not independently reproduced. Evaluation conditions (dataset split, input resolution, IoU aggregation, hardware, precision, timing scope) are not yet documented.
Inference: 43+ FPS; Reported in the previous portfolio record; not independently reproduced. Evaluation conditions (dataset split, input resolution, IoU aggregation, hardware, precision, timing scope) are not yet documented.

DRIVABLE SPACE SEGMENTATION

Problem

Dataset & architecture

Segmentation compare

Evaluation context

Reported results

Problem

Dataset & architecture

Segmentation compare

Evaluation context

Reported results

Related