Principle:Facebookresearch Habitat lab Metric Visualization

Knowledge Sources	Habitat-Lab
Domains	Visualization, Evaluation
Last Updated	2026-02-15 02:00 GMT

Overview

Rendering of agent observations and evaluation metrics into visual outputs (images and videos) for qualitative analysis of agent behavior.

Description

Metric Visualization converts raw sensor observations (RGB, depth, occupancy maps) and evaluation metrics into composite images and videos. This enables qualitative assessment of agent behavior alongside quantitative metrics: seeing what the agent sees, overlaying metric values on frames, and producing episode videos for analysis.

Usage

Use during evaluation to generate visual outputs for debugging, paper figures, or qualitative analysis. Enable via the `video_option` config parameter.

Theoretical Basis

Visualization follows a compositing pipeline:

# Abstract visualization pipeline
for step in episode:
    obs_image = tile_observations(rgb, depth, occupancy_map)
    metric_overlay = render_metrics_as_text(metrics)
    frame = composite(obs_image, metric_overlay)
    frames.append(frame)
video = encode_frames(frames, fps=10)

Related Pages

Implemented By

• Implementation:Facebookresearch_Habitat_lab_Observations_to_image