Submitted by Wenbo Ma.

Submission data

Full name	FoundationStereo (TensorRT) with Photometric Harmonization
Description	We run NVLabs FoundationStereo on rectified low-res two-view pairs using a TensorRT engine. Before inference, we apply a lightweight photometric harmonization to reduce exposure/white-balance mismatch (percentile-based tone alignment with gray-world white balance). The network runs at 480×640 and outputs pixel disparity. Disparity is resized back to the original image size with proper horizontal scaling and saved as 32-bit PFM (NaN for invalid). No post-smoothing or left–right consistency filtering is used.
Parameters	Input size (network): 480×640 Enhancement: harmonize (percentile tone alignment + gray-world WB) Disparity: pixel disparity Upscale rule: bilinear resize + multiply by (W_full / W_in) Output format: disp0.pfm (float32, little-endian), NaN for invalid pixels Hardware: NVIDIA Jetson AGX Orin 32 GB, Ubuntu 22.04 Runtime: TensorRT 8.x, CUDA 11.x, Python 3.10, OpenCV 4.x
Publication title	FS-TRT-Harmonize: Photometric Harmonization for ETH3D Low-Res Two-View
Publication authors	1853532
Publication venue	Technical report (online)
Publication URL	https://github.com/1853532/fs-trt-harmonize#technical-report
Programming language(s)	Python 3.10; CUDA 11.x / TensorRT 8.x; C++ (ETH3D evaluator)
Hardware	NVIDIA Jetson AGX Orin 32 GB (Ampere), Ubuntu 22.04, CUDA 11.x, TensorRT 8.x
Website	https://github.com/1853532/fs-trt-harmonize
Source code or download URL	https://github.com/1853532/fs-trt-harmonize
Submission creation date	15 Aug, 2025
Last edited	15 Aug, 2025

High-res multi-view results

Low-res many-view results

Low-res two-view results

SLAM results