Outdoor Traversability Lab

Arnav Malani · perception → uncertainty → candidate ground Upload an outdoor scene to compare a pretrained semantic baseline with a conservative terrain policy. Green: candidate ground · amber: rough terrain · red: obstacle · gray: unknown.

Outdoor photograph

Minimum semantic confidence

0 1

Maximum normalized entropy

0 1

Candidate boundary margin (image pixels)

0 12

Require connection to bottom-center image region

Baseline: semantic groups

After uncertainty and connectivity filtering

Normalized semantic entropy: red = higher uncertainty

Run measurements

Download mask (IDs 0–3) and report

Video analysis

Upload a clip or load the sample below. The confidence, entropy, margin and connectivity controls above also apply to video. Each sampled frame gets fresh inference; this is not a prerecorded result or a temporal tracking model. Maximum 100 MB, first 20 seconds, and 30 analyzed frames per run.

Outdoor video

Input and filtered terrain (sampled playback)

Sample: Kodaikanal forest walk — Mathanprasath K, source, CC BY 4.0. First 12 seconds, resized, audio removed. Sample-derived overlays retain CC BY 4.0. This is illustrative dense-vegetation footage, not a labeled robotics benchmark.

Analyze first N seconds

1 20

Analyzed frames per second (max 30 frames)

1 2

Video measurements and timestamps

Download comparison MP4, report and mask ZIP

Interpretation

This is an image-space research prototype, using NVIDIA SegFormer-B0 pretrained on ADE20K, without outdoor fine-tuning. Candidate ground is a semantic hypothesis, not a safe driving command. Grass, earth and sand remain caution terrain. Uncertainty scores are uncalibrated; shadows, water, slopes and domain shift can produce confident errors. Connectivity and pixel margins do not measure physical clearance. No depth, traction or robot geometry is inferred.

The project contributes the terrain policy, abstention/connectivity processing, interactive analysis and evaluation harness. Model weights are credited to their original authors. Use a separate, manually labeled outdoor test set before making performance claims.