Evaluation Inspector The Sterkfontein Caves Dataset

Version D0.9-E0.9 — dataset release 0.9, evaluation protocol E0.9

Companion page to The Sterkfontein Caves Dataset (ECCV 2026): ten cave scenes captured in near-darkness with a smartphone. Each method is trained per scene and renders one held-out evaluation view; this page shows every method's saved render beside that view's ground-truth reference, under evaluation protocols you can switch: qualitative results · quantitative results.

The reference is a merge of a ~100-frame burst captured at the evaluation view's camera pose — an HDR+-style burst merge that significantly reduces noise compared to any single frame. References come from dataset release 0.9.

raw-nerfacto and raw-splatfacto are the paper's methods; the others are baselines as run for the paper (references at the bottom). A raw chip on a card means the method reconstructs the scene in the sensor's native raw color space; srgb means it works on developed 8-bit images.

Controls

Qualitative results

D01 Deposit1

low light · ⛶ opens fullscreen at full resolution

reference d01
reference (burst merge)
raw-nerfacto render, d01render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d01render
raw-splatfactoraw
PSNR
SSIM
LPIPS-1K
rawnerf render, d01render
rawnerfraw
PSNR
SSIM
LPIPS-1K
hdrsplat render, d01render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
le3d render, d01render
le3draw
PSNR
SSIM
LPIPS-1K
nerfacto render, d01render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d01render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d01render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d01render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d01render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D02 Deposit2

low light + collocated light · ⛶ opens fullscreen at full resolution

reference d02
reference (burst merge)
raw-nerfacto render, d02render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d02render
raw-splatfactoraw
PSNR
SSIM
LPIPS-1K
rawnerf render, d02render
rawnerfraw
PSNR
SSIM
LPIPS-1K
hdrsplat render, d02render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
le3d render, d02render
le3draw
PSNR
SSIM
LPIPS-1K
nerfacto render, d02render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d02render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d02render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d02render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d02render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D03 Deposit3

spot light · ⛶ opens fullscreen at full resolution

reference d03
reference (burst merge)
raw-nerfacto render, d03render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d03render
raw-splatfactoraw
PSNR
SSIM
LPIPS-1K
rawnerf render, d03render
rawnerfrawdiverged
PSNR
SSIM
LPIPS-1K
hdrsplat render, d03render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
le3d render, d03render
le3draw
PSNR
SSIM
LPIPS-1K
nerfacto render, d03render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d03render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d03render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d03render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d03render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D04 Deposit4

spot light + collocated light · ⛶ opens fullscreen at full resolution

reference d04
reference (burst merge)
raw-nerfacto render, d04render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d04render
raw-splatfactoraw
PSNR
SSIM
LPIPS-1K
rawnerf render, d04render
rawnerfraw
PSNR
SSIM
LPIPS-1K
hdrsplat render, d04render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
le3d render, d04render
le3draw
PSNR
SSIM
LPIPS-1K
nerfacto render, d04render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d04render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d04render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d04render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d04render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D05 Tunnel1

extreme low light · ⛶ opens fullscreen at full resolution

reference d05
reference (burst merge)
raw-nerfacto render, d05render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d05render
raw-splatfactorawdiverged
PSNR
SSIM
LPIPS-1K
rawnerf render, d05render
rawnerfraw
PSNR
SSIM
LPIPS-1K
hdrsplat render, d05render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
le3d render, d05render
le3draw
PSNR
SSIM
LPIPS-1K
nerfacto render, d05render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d05render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d05render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d05render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d05render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D06 Tunnel2

extreme low light + artificial light · ⛶ opens fullscreen at full resolution

reference d06
reference (burst merge)
raw-nerfacto render, d06render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d06render
raw-splatfactoraw
PSNR
SSIM
LPIPS-1K
rawnerf render, d06render
rawnerfraw
PSNR
SSIM
LPIPS-1K
hdrsplat render, d06render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
le3d render, d06render
le3draw
PSNR
SSIM
LPIPS-1K
nerfacto render, d06render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d06render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d06render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d06render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d06render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D07 Protrusion1

natural light · ⛶ opens fullscreen at full resolution

reference d07
reference (burst merge)
raw-nerfacto render, d07render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d07render
raw-splatfactoraw
PSNR
SSIM
LPIPS-1K
rawnerf render, d07render
rawnerfraw
PSNR
SSIM
LPIPS-1K
hdrsplat render, d07render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
no saved render
le3d
nerfacto render, d07render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d07render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d07render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d07render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d07render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D08 Protrusion2

natural + artificial light · ⛶ opens fullscreen at full resolution

reference d08
reference (burst merge)
raw-nerfacto render, d08render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d08render
raw-splatfactorawdiverged-exposure
PSNR
SSIM
LPIPS-1K
rawnerf render, d08render
rawnerfrawwatch-superposition
PSNR
SSIM
LPIPS-1K
hdrsplat render, d08render
hdrsplatrawsuperposition
PSNR
SSIM
LPIPS-1K
le3d render, d08render
le3drawsuperposition
PSNR
SSIM
LPIPS-1K
nerfacto render, d08render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d08render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d08render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d08render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d08render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D09 RoofPendant1

slow shutter · ⛶ opens fullscreen at full resolution

reference d09
reference (burst merge)
raw-nerfacto render, d09render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d09render
raw-splatfactoraw
PSNR
SSIM
LPIPS-1K
rawnerf render, d09render
rawnerfraw
PSNR
SSIM
LPIPS-1K
hdrsplat render, d09render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
le3d render, d09render
le3draw
PSNR
SSIM
LPIPS-1K
nerfacto render, d09render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d09render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d09render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d09render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d09render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

D10 RoofPendant2

fast shutter · ⛶ opens fullscreen at full resolution

reference d10
reference (burst merge)
raw-nerfacto render, d10render
raw-nerfactoraw
PSNR
SSIM
LPIPS-1K
raw-splatfacto render, d10render
raw-splatfactoraw
PSNR
SSIM
LPIPS-1K
rawnerf render, d10render
rawnerfraw
PSNR
SSIM
LPIPS-1K
hdrsplat render, d10render
hdrsplatraw
PSNR
SSIM
LPIPS-1K
no saved render
le3d
nerfacto render, d10render
nerfactosrgb
PSNR
SSIM
LPIPS-1K
splatfacto render, d10render
splatfactosrgb
PSNR
SSIM
LPIPS-1K
3dgs_dk render, d10render
3dgs_dksrgb
PSNR
SSIM
LPIPS-1K
aleth_nerf render, d10render
aleth_nerfsrgbenc-ambiguous
PSNR
SSIM
LPIPS-1K
luminance-gs render, d10render
luminance-gssrgb
PSNR
SSIM
LPIPS-1K

Quantitative results

Renders are scored under the calibrated protocol, which brings each render into the reference's geometry and color space, and under two fitted radiometric variants that we keep separate because they forgive different things: cal + WB allows only what a physical capture error can produce (a per-channel white-balance gain and black-level offset), while cal + affine is an unconstrained per-channel gain and offset that can also absorb color differences no camera setting explains. The gap between the two measures exactly that extra forgiveness. A raw-space evaluation further below compares the raw-output methods directly against the sensor measurements, with no development or exposure normalization. One caveat throughout: a reconstruction that integrates many views can be lower-noise than the reference itself, and pixel metrics then penalize the render for not reproducing the reference's residual noise — take the numbers with a pinch of salt 🧂 (see d07 raw-nerfacto in fullscreen: the render's shadows are cleaner than the reference's).

Linear sRGB evaluation

protocol details
methodmetricd01d02d03d04d05d06d07d08d09d10mean
raw-nerfactoPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
raw-splatfactoPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
rawnerfPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
hdrsplatPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
le3dPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
nerfactoPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
splatfactoPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
3dgs_dkPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
aleth_nerfPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓
luminance-gsPSNR↑SSIM↑LPIPS-1K↓LPIPS-4K↓

Flags (grey cells and card chips) — diverged: the training run collapsed. diverged-exposure: trained, but at a wildly wrong brightness. superposition / watch-superposition: the d08 model was trained with mixed image orientations and contains both (the d08 evaluation view was captured with the phone physically inverted). enc-ambiguous: the method's output encoding is inconsistent across scenes. no saved render: not present in the paper's result archive. Flagged pairs remain in the means; missing pairs are excluded.

Raw-space evaluation

Raw-output methods only (the five with a raw chip); protocol-independent. The float32 raw render is compared directly against the reference burst-merge mosaic, sampled at each pixel's color-filter-array (CFA/Bayer) channel, in DN. A DN (digital number) is one integer step of the sensor's analog-to-digital output — the native unit of the raw files; these are 10-bit, so values run 0–1023 with white level 1023. No development pipeline and no exposure normalization touches either side; d08's inverted capture is handled as in the calibrated protocol.

methodmetricd01d02d03d04d05d06d07d08d09d10mean
raw-nerfactoL1 (DN)↓rel err↓exposure ×0.8817.4%0.960.7820.8%1.0920.6019.7%0.8519.9578.7%1.141.9158.9%1.222.9083.7%1.401.7435.6%1.534.4458.0%1.6243.7292.7%1.643.9356.3%1.5810.0852.2%1.30
raw-splatfactoL1 (DN)↓rel err↓exposure ×1.0519.2%0.871.0136.8%1.0418.8825.2%0.9724.05165.8%0.946.24225.7%1.472.7463.1%0.522.2485.2%1.2916.64808.9%0.9333.81112.6%1.383.7568.6%0.8411.04161.1%1.03
rawnerfL1 (DN)↓rel err↓exposure ×0.9817.9%0.870.6518.1%1.0188.4380.0%0.1418.2738.7%0.841.3637.4%1.102.0646.0%1.211.7440.2%1.384.2761.4%0.9437.2670.0%0.804.0045.6%0.4015.9045.5%0.87
hdrsplatL1 (DN)↓rel err↓exposure ×1.2526.0%0.820.8929.1%0.9916.9316.6%0.979.4422.5%0.951.2838.4%0.811.3737.1%0.851.9662.5%1.303.2159.9%0.9525.8770.8%1.192.2443.0%0.886.4440.6%0.97
le3dL1 (DN)↓rel err↓exposure ×1.1621.7%0.900.7922.8%0.8719.0119.0%0.9415.4232.8%0.870.9526.8%0.901.2329.8%1.003.2481.4%1.0126.5433.3%0.658.5433.4%0.89

exposure × = median render/reference over pixels ≥5 DN (1.00 = absolute photometric agreement). L1 = mean |error| in DN. rel err = mean |error| / max(reference, 1 DN). Raw PSNR is not reported: measured against the full 0–1023 range it comes out above 55 dB for every method — not because the reconstructions are near-perfect, but because these scenes occupy only the bottom few percent of the range — and it does not compare across scenes of different brightness.

Downloads

Full-resolution images: eval-inspector assets D0.9-E0.9 (paper-run renders scored against dataset 0.9 references under evaluation protocol E0.9 — the "calibrated" protocol on this page), served from the dataset's public Globus collection. Asset versions name the (dataset release, evaluation protocol) pair and are immutable once published. Protocol versions are defined in the code repository's EVAL_PROTOCOL.md; E0.9 is current, and E1.0 is reserved for the declared-stable protocol.

The float32 raw-space method outputs — the exact files scored here, suitable for new evaluation protocols — and the dataset itself are distributed through the dataset's Globus collection; see the project page for access.

References