nv-dvl

nv-dvl / capa

Public

CAPA: Depth Completion as Parameter-Efficient Test-Time Adaptation

11
0
100% credibility
Found Apr 14, 2026 at 11 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

CAPA enhances sparse depth maps from monocular depth models by efficiently adapting them at test time using techniques like LoRA and visual prompt tuning.

How It Works

1
🔍 Discover CAPA

You hear about CAPA, a simple way to turn blurry, spotty depth maps from your photos or videos into clear, full ones.

2
📥 Grab sample scenes

Download ready example images or short videos with partial depth info to try it out.

3
⚙️ Pick your favorite setup

Choose one of the easy presets, like the stable visual prompt option for smooth results.

4
🚀 Run depth completion

Hit go, and watch as it smartly fills in the missing depths using a quick learning trick tailored to your scene.

5
👀 See the magic

Enjoy colorful depth views and side-by-side comparisons showing the before-and-after improvement.

✅ Perfect depths ready

You now have sharp, complete depth maps for your images or videos, ready for 3D projects or fun experiments.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 11 to 11 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is capa?

CAPA turns sparse depth inputs—like noisy LiDAR scans paired with RGB images—into dense, accurate depth maps via test-time adaptation of SOTA monocular depth models (VGGT, MoGe-2, UniDepth-v2). You feed it .pt files with RGB frames, sparse depth, and masks via a simple CLI (`python run.py --config vggt_vpt.yaml --input data/`), and it optimizes lightweight LoRA or VPT adapters on-the-fly per scene, outputting completed depths and visualizations. Built in Python, it handles single images or video sequences, aligning predictions to conditions for metric-scale results.

Why is it gaining traction?

Unlike rigid zero-shot depth estimators, CAPA adapts frozen models parameter-efficiently at inference, boosting accuracy on sparse real-world data without full finetuning. VPT configs offer reproducible results, and separate envs sidestep dep conflicts across models. As a capa tool github download, its NVIDIA-backed paper and sample data make test-time adaptation accessible for quick experiments.

Who should use this?

Robotics engineers fusing LiDAR with cameras for navigation, CV researchers benchmarking depth completion on datasets like ScanNet or IBims, or AR devs needing per-scene metric depths from sparse sensors. Ideal if you're evaluating github capa rules for adaptation in low-data regimes.

Verdict

Promising research prototype for depth completion—grab capa download github for proofs-of-concept, but 1.0% credibility reflects 11 stars and non-commercial license limiting production use. Docs are solid with configs and samples; pair with flare capa github for rules if extending.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.