Room Envelopes: Structural, Acoustic & Thermal Aspects

• Room envelopes are task-dependent boundaries that can represent structural layouts, acoustic decay, or thermal partitions in indoor environments.
• They apply in diverse fields such as vision-based reconstruction, robotics mapping, acoustic impulse response modeling, and building physics for energy efficiency.
• Recent methods transition from global topology prediction to dense per-pixel and occupancy-grounded techniques, enhancing geometry recovery and performance metrics.

“Room envelopes” denotes a family of related constructs rather than a single canonical object. In indoor scene understanding, the term usually refers to the structural layout of a room—walls, floors, ceilings, windows, and doors—either as the visible image-plane projection of that layout or as the first structural surface that would remain if furniture and fixtures were removed. In room acoustics, by contrast, an envelope is the slowly varying decay structure of late reverberation. In building physics, the envelope is the wall assembly whose composition controls heat transfer and passive comfort. The term is therefore unified less by a single representation than by a common role: it names the boundary structure through which a room is modeled, inferred, or optimized in a task-specific way (Bahrami et al., 6 Nov 2025, Lee et al., 2017, Lin et al., 2024, Bojic et al., 2013).

1. Terminological scope and representational families

The literature uses “room envelope” in several technically distinct senses. In image-based indoor reconstruction, it typically denotes the structural layout behind clutter. In occupancy-based robotics, it denotes a polygonal free-space footprint attached to a room node. In acoustics, it denotes decay envelopes governing room impulse responses. In building engineering, it denotes the wall-layer composition that mediates heat storage and attenuation (Bahrami et al., 6 Nov 2025, Lee et al., 2017, Lin et al., 2024, Bojic et al., 2013, Zumaya et al., 11 Jun 2026).

Research area	Meaning of “room envelope”	Typical representation
Monocular layout estimation	Visible image-plane projection of walls, floor, and ceiling	Ordered keypoints plus room type
Indoor structural reconstruction	First visible surface and first structural layout surface	RGB with visible/layout pointmaps
Occupancy-grounded scene graphs	Tracked free-space support of a room	2D polygonal footprint
Room acoustics	Slowly varying late-reverberation decay structure	Multi-exponential subband envelopes
Building thermal design	Heat-transferring room boundary assembly	Wall-layer thicknesses

This plurality matters because methods are not directly interchangeable across these meanings. A pointmap predictor for structural layout, a polygonal free-space decomposer, and a multi-exponential reverberation model all describe “room envelopes,” but they do so at different physical levels and under different observability assumptions.

2. Image-plane and pointmap room envelopes

A major computer-vision formulation treats room envelopes as a constrained 2D layout problem. RoomNet formulates room layout estimation as recovery of a “2D boxy room envelope” from a single RGB image under the Manhattan world assumption. Instead of segmentation followed by geometric hypothesis ranking, it predicts an ordered set of room-layout keypoints together with a room-type label. The representation uses 11 layout/topology types indexed from 0 to 10 and 48 total keypoints; once the correct type is known, the corresponding keypoints are connected in predefined order to recover the polygonal room boundary and the induced wall/floor/ceiling segmentation. The network is an encoder-decoder CNN inspired by SegNet, takes $320 \times 320$ RGB input, produces 48 heatmaps at $40 \times 40$ resolution, uses Gaussian supervision with standard deviation 5 pixels, sets the classification weight to $\lambda = 5$, and downweights background gradients by a factor of 0.2. On Hedau, RoomNet recurrent 3-iter reaches 8.36% pixel error; on LSUN, it reaches 6.30% keypoint error and 9.86% pixel error, while running at 5.96 FPS and about 200× faster than Dasgupta et al. (Lee et al., 2017).

The more recent “Room Envelopes” dataset shifts the representation from ordered 2D keypoints to pixel-aligned 3D supervision. It is built from Hypersim and contains 77,400 images across 461 indoor scenes. Each sample provides an RGB image, a visible-surface pointmap, and a layout-surface pointmap, together with metric-scale scene point clouds, camera information, and per-pixel normals. The layout target is defined as the first structural surface visible along each ray if fittings and fixtures are removed, and the paper explicitly describes the structural layout as comprising walls, floors, ceilings, windows, and doors. The split is by scene, 365 / 46 / 46, and unseen layout pixels comprise 23.57% of all layout pixels in the test set. Empirically, fine-tuning MoGe v1 on this supervision improves prediction precisely where ordinary monocular geometry is weakest: on unseen layout regions, Chamfer Distance improves from 1.037 to 0.753 relative to pre-trained MoGe, while $[email protected]$ rises from 0.114 to 0.512 and $[email protected]$ from 0.059 to 0.390. The improvement is not uniform across all metrics: on seen layout pixels, pre-trained MoGe retains the best Chamfer Distance, 0.229 versus 0.362 for the layout-trained model, but the layout-trained model attains the best overall F-scores, 0.633 at 0.1 and 0.499 at 0.05 (Bahrami et al., 6 Nov 2025).

Taken together, these two lines of work illustrate a transition from topological envelope recovery under fixed boxy layouts to dense per-pixel 3D layout surfaces. The former enforces global structural coherence through topology classes; the latter emphasizes amodal structural geometry under occlusion.

3. Occupancy-grounded polygonal room footprints

A distinct robotics-oriented usage defines room envelopes as occupancy-grounded polygonal extents. In an occupancy-grounded hierarchical 3D scene graph, each room node is anchored to a single tracked free-space region, giving the room an explicit, persistent geometric footprint defined by observed free space. The construction begins from an OctoMap-based 3D occupancy map at 0.05 m voxel resolution and 8 m sensor integration range, then collapses it into a 2D free-space layer by checking whether each vertical voxel column contains a free run of at least $R_{\max Z}=0.3$ m. That 2D layer is decomposed into polygonal regions with incremental DUDE, and those polygons are tracked over time using intersection-over-union, centroid displacement, and area-ratio gating (Zumaya et al., 11 Jun 2026).

The tracking thresholds are explicit. Association requires IoU $> 0.20$, centroid displacement $< 1.5$ m, and area ratio in $[0.25, 4.0]$. Unmatched regions are retained for up to 3 consecutive updates. Entire decompositions are rejected if total free area or region count drops by more than 15% relative to the previous valid decomposition. A room-maintenance pass runs at 0.5 Hz, and each tracked region supports at most one room. The resulting room envelope is therefore not a full 3D room solid but a 2D ground-plane polygon derived from occupancy-informed free-space structure.

Evaluation is also polygonal. On 12 Matterport3D scenes with 4 to 23 ground-truth rooms, predicted and annotated room polygons are centroid-aligned, matched one-to-one by Hungarian assignment, and accepted when $\mathrm{IoU}\geq 0.15$. Boundary F1 uses 0.25 m tolerance and 0.1 m boundary sampling. Relative to Hydra, the occupancy-grounded method raises recall from 0.152 to 0.379 and F1 from 0.241 to 0.427, while precision drops from 0.703 to 0.518. Matched-overlap and boundary quality remain modest for both approaches, with mIoU 0.364 versus 0.323 and bF1 0.224 versus 0.113. The paper’s explicit conclusion is that occupancy-grounded anchoring recovers substantially more room instances than place-connectivity construction, but wall-accurate room boundaries remain an open problem for both methods (Zumaya et al., 11 Jun 2026).

This definition of room envelope is narrower than full floorplan reconstruction and more geometric than place clustering. It is well suited to containment, adjacency, and room-scale reasoning, but it inherits errors from free-space decomposition: if a region merges two physical rooms, the room node merges them; if the decomposition fragments one room, the room layer fragments as well.

4. Acoustic room envelopes and latent room descriptors

In acoustics, the envelope is not the geometric boundary itself but a compact representation of how that boundary shapes sound. DECOR, introduced for “RIR completion,” assumes that direct sound and early reflections encode enough information about room geometry and absorption to constrain the late tail of a room impulse response. The task is

$40 \times 40$0

so the input is the first 50 ms of the RIR and the target is the remaining 950 ms. The late field is modeled as filtered noise under a multi-exponential decay envelope: $40 \times 40$1 DECOR extends this to $40 \times 40$2 subbands of filtered noise, uses $40 \times 40$3 decay times linearly sampled from 0.05 to 3.0 s, sets the FIR filter order to $40 \times 40$4, maps a 50 ms head $40 \times 40$5 to a latent vector of dimension $40 \times 40$6, and predicts amplitudes through

$40 \times 40$7

Training uses a multiresolution STFT loss rather than direct envelope supervision, while evaluation includes EDF error, $40 \times 40$8 error, and DRR error. On the test set, DECOR reports MSTFT 1.073, EDF MAE 4.187 dB, EDF RMSE 5.77 dB, $40 \times 40$9 MSE 0.0538 s, and DRR MSE 1.106 dB, remaining comparable to an adapted FiNS baseline while producing smoother reverberation (Lin et al., 2024).

Adjacent work shifts from explicit decay envelopes to latent room signatures. RevRIR learns a joint embedding of RIRs and reverberant speech for room fingerprinting and room-shape classification. Its simulated dataset contains 110 room classes: 16 small rooms, 52 large rooms, and 42 halls, parameterized by width, depth, and height. The best speech-side model attains 60% Top-1 accuracy on 110-room classification, while the frozen speech encoder reaches 95.4% accuracy on the 3-room-type task. The reported t-SNE organization is speaker-independent and content-independent, with visible semantic ordering by width and depth, though not height (Bitterman et al., 2024). A further development learns a structured room-acoustic latent space from 3000 measured RIRs, aligns a 4096-dimensional speech embedding to that latent space, and calibrates a scalar uncertainty score $\lambda = 5$0 using corruption-induced dispersion

$\lambda = 5$1

That framework reports AP 0.99 for RIR verification and Spearman correlation 0.90 between uncertainty and dispersion (Xiang et al., 1 Jul 2026).

This suggests a broader acoustic interpretation of “room envelope”: not only the explicit multi-exponential decay law of late reverberation, but also a latent room-dependent descriptor anchored to RIR structure. In that sense, acoustic room envelopes are indirect, transfer-based surrogates for enclosure geometry rather than geometric reconstructions of walls and corners.

5. Building envelopes as thermal room boundaries

In building physics, the room envelope is the material assembly that controls indoor heat transfer and passive comfort. A representative study on Reunion Island models a naturally ventilated one-storey residential house with floor area $\lambda = 5$2 and five thermal zones: two bedrooms, a kitchen, a living/main room, and a toilet. The wall assembly is a three-layer external wall customary in Reunion construction, ordered from outside to inside as wood, glass wool insulation, and concrete block. Reported nominal baseline thicknesses are 25 mm wood, 25 mm glass wool, and about 203.2 mm concrete block, although the results section later describes the baseline house as 160 mm concrete block, 25 mm insulation, and 25 mm wood, an inconsistency explicitly noted in the reproduced details (Bojic et al., 2013).

The optimization is performed one layer at a time with EnergyPlus, GenOpt, and the Hooke–Jeeves method. The design variable is the thickness of the selected layer, constrained to

$\lambda = 5$3

Comfort is evaluated using a weighted annual Fanger PMV objective,

$\lambda = 5$4

with weights $\lambda = 5$5, $\lambda = 5$6, $\lambda = 5$7, and $\lambda = 5$8. The study describes this as maximizing comfort, but the reported formulation and discussion show that it minimizes $\lambda = 5$9, moving positive PMV values toward neutral.

The baseline house has total PMV about 1.45. The best-performing optimization thickens the concrete block from 160 mm to 200 mm, leaving wood and insulation at 25 mm each and raising total wall thickness from 210 mm to 250 mm. That reduces total PMV to about 0.55. Thickening the wood layer to 100 mm also brings total PMV to around 0.5, but the paper notes that such a wood layer is unrealistic because wood is normally decorative. Thickening glass wool to 100 mm is thermally beneficial but least favorable among the three optimized options; the kitchen PMV remains about 1.17, and the insulation thickness is increased by 300%. The study’s explicit design message is that, for this hot-humid naturally ventilated case, moderate increase of concrete-block thickness is more effective and practical than very thick insulation or decorative wood (Bojic et al., 2013).

Here the room envelope is not an inferred scene boundary but a heat-transferring material boundary. Its optimization depends on the balance between resistance and thermal mass, and the paper argues that tropical comfort is governed not only by steady-state resistance but also by transient heat storage and damping.

6. Common assumptions, limits, and open problems

Across these literatures, room-envelope models are strongly shaped by what is treated as observable and what is treated as latent. Image-plane layout estimation in RoomNet is confined to 11 predefined boxy room types under the Manhattan world assumption, and the method is reported to fail when room boundaries are barely visible or when multiple plausible layouts exist (Lee et al., 2017). The Room Envelopes dataset removes that explicit topology restriction, but its layout pointmaps can still contain holes wherever no view in the multiview capture ever observes the structural surface, and its real-world evidence is qualitative rather than benchmarked quantitatively (Bahrami et al., 6 Nov 2025).

Occupancy-grounded room polygons avoid explicit Manhattan assumptions, but they reduce rooms to 2D footprints rather than full 3D volumes. Because each tracked region supports at most one room, room granularity is inherited directly from free-space decomposition, and the paper identifies merged rooms, local collapse at narrow connections, and reconstruction leakage as characteristic failure modes (Zumaya et al., 11 Jun 2026). Acoustic envelope models make a different trade-off: DECOR does not estimate geometry directly, but instead assumes that late reverberation can be represented as filtered stochastic noise under a sparse multi-exponential envelope and that 50 ms of head is sufficient input; its performance degrades on an unseen dataset, indicating limited generalization under dataset mismatch (Lin et al., 2024). RevRIR and the later uncertainty-calibrated room-embedding framework likewise operate on room-dependent transfer signatures rather than geometric walls, and both are tied to training regimes where room classes or RIR identities structure the latent space (Bitterman et al., 2024, Xiang et al., 1 Jul 2026).

Building-envelope optimization has its own constraints: the Reunion study optimizes one wall layer at a time, fixes infiltration at $[email protected]$0 ach, evaluates comfort with weighted PMV rather than adaptive comfort, and reports numerical inconsistencies in both baseline wall thickness and percentage improvement (Bojic et al., 2013). These are not incidental details; they show that even in the most material sense of “room envelope,” the optimized boundary remains conditional on modeling choices.

This suggests that “room envelope” is best understood as a family of task-dependent abstractions. In vision it is often a structural surrogate for occluded architecture; in robotics it is a polygonal support for room-scale reasoning; in acoustics it is a decay law or latent room signature; in building science it is the thermophysical boundary assembly itself. The shared concept is the room-defining interface, but the mathematically relevant object changes with the question being asked.