BenchmaRking Anisotropic vEM Evaluation for Axial Super-Resolution(BRAVE-ASR)

BRAVE-ASR (Benchmarking Robust Anisotropic Volume-microscopy Evaluation for Axial Super-Resolution) is a benchmark dataset for evaluating z-axis super-resolution methods in volume electron microscopy, introduced as part of the SkelEM framework (ECCV 2026). It contains co-aligned volumes of mouse brain tissue acquired using a Thermo Fisher Plasma-FIB (pFIB) at 5 nm XY pixel size and 1 µs dwell time, with four primary volumes at physical slice thicknesses of 5 nm, 10 nm, 20 nm, and 40 nm (SIFT-registered and histogram-normalized), plus a larger raw 5 nm volume without histogram normalization. A key finding enabled by this dataset is that slice sampling (retaining every N-th slice) is a substantially more faithful simulation of real anisotropic acquisition than average downsampling (mean pooling of N adjacent slices). Through sectional Fourier Shell Correlation and Fréchet Inception Distance analysis against the physically acquired ground truth, we show that average downsampling artificially suppresses noise by a factor of 1/r and attenuates high-frequency content, producing an unrealistically "clean" input that inflates the apparent performance of ASR methods. Slice sampling preserves the true noise statistics and spectral characteristics of real acquisitions across all tested anisotropy ratios (2×, 4×, 8×). We provide an evaluation script to reproduce all results. See the full README for details, usage instructions, and citation information.