EmiState-HDR is a rendered HDR indoor dataset with an evaluation protocol for storing additional observed emissive states of a scene compactly. Each transition takes a scene from a base state to a target state in which a light source (room lighting or a local emitter such as a screen, sign, menu board or picture frame) has changed intensity. A method receives only the images and cameras of 80 base-state and 80 target-state training views, builds a representation of the target state, and is scored at 40 evaluation views on the target state and three intermediate states.

This repository contains the dataset used in the paper EmiState-HDR (H. Lee, J. Ko, J. Lee). Code: coming soon at https://github.com/hogun5501/EmiState-HDR (project page: https://hogun5501.github.io/EmiState-HDR/).

Target state of the 13 configurations (rows: scenes).

A (room light) and B (local emitter) transitions at α = 0, 0.2, 0.5, 0.8, 1; every frame is a separately rendered ground-truth state.


Scenes (geometries)	Cafe, Gallery, Transit, TV
Configurations	13 (Cafe 3, Gallery 3, Transit 3, TV 4)
Transitions	26 (an A and a B transition per configuration)
Cameras per configuration	200
Rendered EXRs	33,000 (165 state folders × 200 views)
Split per transition	80 base-state train / 80 target-state train / 40 evaluation (no overlap)
Real capture	Office (2 states × 120 frames)

Scene	Resolution	Samples	EXR	Compositor post-processing
Cafe	1200×1200	128	half	yes
Gallery	1200×1200	128	half	yes
Transit	1024×1024	128	half	no
TV	1024×1024	1024	float	no

All images are rendered with Blender Cycles. Per-image Blender version, samples, scene file name, render date and precision, read from the EXR headers, are listed in render_settings.csv.

Transitions

In each configuration two base states lead to one shared target state in which the local emitter and the room lighting are both on:

A: the room lighting increases while the local emitter stays on.
B: the local emitter increases while the room lighting stays on.

The room lighting consists of emissive materials in Cafe, Gallery and Transit and of area lights in TV. Gallery and Transit additionally contain an HDRI environment map (background strength 1.0) that is identical in all states; Cafe and TV have no fixed lighting.

State folders are named scA_xx / scB_xx, where xx is the state value α×100 of the changing light source (α = 0 base, α = 1 target). Intermediate states α = 0.2, 0.5, 0.8 exist for all scenes; TV additionally stores every 0.1. The A and B target states are the same render, stored once as scB_100 (TV: scA_100); transitions/<name>.json maps each logical state of a transition to its folder.

Layout

scenes/<scene>/<configuration>/<state>/Camera_###.exr   rendered HDR, linear, as saved by the renderer
scenes/<scene>/<configuration>/cameras.json             intrinsics + camera-to-world poses of the 200 cameras
transitions/<transition>.json                           states (logical state -> folder) and the 80/80/40 view split
masks/<scene>_Camera_###.exr                            emissive-material masks at the evaluation views
audit/                                                  per-view ground-truth audit results and summaries
render_settings.csv                                     per-image render metadata
office/                                                 real Office capture (see below)
tools/emistate_hdr.py                                   reading, display transform, PSNR, compositor inversion

cameras.json follows the NeRF-synthetic (Blender) convention: transform_matrix is camera-to-world with the camera looking along −z and y up. The same file also gives fl_x, fl_y, cx, cy, w, h.

A transition file looks like:

{
 "transition": "cafe_01_A", "scene": "cafe", "configuration": "cafe_01", "path": "A",
 "source_state": "scA_00", "target_state": "scA_100",
 "states": {"scA_00": "scenes/cafe/cafe_01/scA_00", "...": "...", "scA_100": "scenes/cafe/cafe_01/scB_100"},
 "train_views_source_state": ["Camera_###", "..."],
 "train_views_target_state": ["Camera_###", "..."],
 "test_views": ["Camera_###", "..."]
}

Radiometry

The EXRs hold unclipped HDR values exactly as saved by the renderer. The float32 arrays used for evaluation in the paper are these EXR values without any change (verified for every array).

In the Cafe and Gallery renders the scene compositor blended the pre- and post-denoising images 1:1 and applied a brightness-dependent Color Correction before saving. tools/emistate_hdr.py: invert_compositor inverts the brightness-dependent scaling per pixel. Because the denoised half of the blend is not linear in the light intensities, the inverted values are only approximately linear. Transit and TV have no post-processing. In the paper, methods are trained and evaluated on the EXR values as stored; the audit and the surround measurement use the inverted values.

Evaluation uses the display transform f(x) = x / (1 + x) on both prediction and ground truth, with PSNR over the full image and RGB.

from tools.emistate_hdr import load_transition, image_path, read_exr, psnr
s = load_transition('.', 'cafe_01_A')
gt = read_exr(image_path('.', s, s['target_state'], s['test_views'][0]))

Masks

Emissive-material masks are rendered at the evaluation views with every material replaced by an emission shader, 1 sample per pixel and a pixel filter width of 0.01. Emissive pixels have G = 0.5 and R = material ID × 0.01; a pixel belongs to the mask when G > 0.25. Reflections of emitters are not part of the mask.

Audit and known deviations

Every intermediate state is checked per view against the linear relation between the two endpoints (audit/). Of 25,200 audited views, 303 exceed the thresholds; they are kept and listed. Most of them (278) are in the B-transition intermediate states of TV natural, whose target state was rendered from a different Blender scene file than its intermediate states; the screen brightness of those intermediate states differs from the linear prediction by up to about 5% depending on the view. Target-state evaluation is not affected.

Office (real capture)

office/office_012_013/ holds a real two-state capture of an office with the room lighting on: three table lamps off (off) and on (on). Each state was recorded separately with the front lens of an Insta360 X5 along its own trajectory, so there are no same-view pairs. Per state, 120 frames are split into 100 training and 20 evaluation images (*_train, *_test; *_full holds all 120). The 240 frames were registered jointly with COLMAP; images/ contains the undistorted PINHOLE images (2052×923) of that registration and sparse/0 the matching model. raw_frames/ contains the 1600×900 frames before undistortion. The observations are LDR JPEGs.

License and assets

The data are released under CC BY-NC 4.0. The scene assets come from BlenderKit, whose terms do not allow redistributing asset files outside a project; therefore only rendered results are released, not the Blender scenes. The images shown on the local emitters were created by the dataset author with generative AI and then edited, except the TV natural screen image, which is a free image usable without copyright restrictions.

Citation

@misc{emistatehdr,
  title  = {EmiState-HDR},
  author = {Lee, Hogeon and Ko, Jaerin and Lee, Jangwon},
  year   = {2026},
  doi    = {10.57967/hf/10897},
  url    = {https://huggingface.co/datasets/logun99/EmiState-HDR}
}

Downloads last month: 86

Total file size:

272 GB

EmiState-HDR

Contents