Datasets:

Botya
/

FH6-Temporal-Upscaling-Benchmark

license: cc-by-nc-4.0
task_categories:
  - image-to-image
  - video-classification
tags:
  - super-resolution
  - temporal-upscaling
  - amd-fsr
  - taa
  - computer-vision
  - video-restoration
  - benchmark
pretty_name: Forza Horizon Temporal Upscaling Benchmark (TAA vs. FSR 4.1.1)
size_categories:
  - 1K<n<10K

Forza Horizon Temporal Upscaling Benchmark (TAA vs. AMD FSR 4.1.1)

Dataset Summary

This dataset provides 1,041 pixel-aligned, lossless 512x512 RGB paired image tensors (2,082 total patches) curated to evaluate temporal upscaling artifacts—focusing on disocclusion ghosting, high-frequency specular shimmering, and motion smearing—comparing AMD FidelityFX Super Resolution (FSR 4.1.1) against Native Temporal Anti-Aliasing (TAA) ground truth.

Each sample includes horizontal (x_shift_pixels) and vertical (y_shift_pixels) sub-pixel temporal drift metrics computed via 2D Fourier Phase Correlation, validating strict spatial alignment without applying synthetic affine warping or lossy interpolation to the raw pixel data.

Capture & Synchronization Methodology

Benchmark Hardware & Rendering Pipeline

Hardware Architecture: AMD Ryzen AI 9 / Radeon 890M integrated graphics architecture.
Display Output: 1920x1080 native rasterization; engine V-Sync disabled to prevent buffered frame-queue latency.
Frame Pacing Lock: Hardware-level frametime limiter enforced via RivaTuner Statistics Server (RTSS) locked to 30.00 FPS (strict 33.33 ms frame delivery, zero frame drops).
Capture Format: OBS Studio AMF hardware recording to Constant Frame Rate (CFR) Matroska (.mkv) containers utilizing 4:4:4 planar color sampling (I444), eliminating chroma subsampling artifacts (4:2:0).
Patch Extraction: Lossless decoding to 24-bit RGB tensors via OpenCV. Central 512x512 regions were cropped to isolate vehicle motion and reflection vectors while excluding all 2D HUD text, UI alerts, and physics debris.

Sub-Pixel Verification Telemetry

Temporal synchronization was achieved through temporal Mean Squared Error (MSE) minimization over a HUD-isolated search window. Exact sub-pixel spatial displacement was logged per frame using 2D Fourier Phase Correlation:

$\Delta x, \Delta y = \arg\max_{(x,y)} \mathcal{F}^{-1} \left\{ \frac{\mathcal{F}(I_{\text{TAA}}) \cdot \mathcal{F}^*(I_{\text{FSR}})}{\vert{}\mathcal{F}(I_{\text{TAA}}) \cdot \mathcal{F}^*(I_{\text{FSR}})\vert{}} \right\}$

Dataset Structure & Schema

The dataset contains a single train split consisting of 1,041 verified image pairs:

Feature Name	Type	Description
`clip_id`	`string`	Anchor identifier of the source video window.
`frame_index`	`int32`	Sequential frame index within the extracted window (`1`–`150`).
`frame_filename`	`string`	Source frame filename (`frame_00001.png` to `frame_00150.png`).
`taa_ground_truth`	`Image`	512x512 lossless PNG patch rendered with Native TAA.
`fsr_upscaled`	`Image`	512x512 lossless PNG patch rendered with AMD FSR 4.1.1.
`x_shift_pixels`	`float32`	Horizontal sub-pixel displacement calculated via Phase Correlation.
`y_shift_pixels`	`float32`	Vertical sub-pixel displacement calculated via Phase Correlation.

Quickstart Usage

Stream or download the dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

# Stream the dataset directly into Python
dataset = load_dataset("Botya/FH6-Temporal-Upscaling-Benchmark", split="train")

# Retrieve and inspect the first paired sample
sample = dataset[0]
taa_patch = sample["taa_ground_truth"]
fsr_patch = sample["fsr_upscaled"]

print(f"Loaded {sample['clip_id']} [{sample['frame_filename']}]")
print(f"Spatial Drift: dX={sample['x_shift_pixels']:.4f}px, dY={sample['y_shift_pixels']:.4f}px")