# Tutorial image provenance `spherical-object-hypothesis-method.svg` and `application-strategy-routes.svg` are original PanorAi vector diagrams derived only from the documented geometry, execution-strategy, and object-localization contracts. They contain no third-party image or generated model output. The photographic tutorial assets come from the public Poly Haven HDRIs published under **CC0**. ## Spherical region-analysis diagrams The six `region-analysis-*.svg` files are original PanorAi scientific figures generated by `scripts/generate_region_analysis_figures.py`. Each output is computed from an analytic ERP fixture with the public region-analysis API; no external image, dataset, or third-party media is used. The figures deliberately show longitude periodicity, support and validity, solid angle, angular distance, bearings, and adjacency instead of presenting planar mock-ups. Regenerate them from the repository root with: ```bash python scripts/generate_region_analysis_figures.py \ --output-dir docs/_static/tutorials ``` The six `region-analysis-real-*.jpg` files are real-image integration evidence generated by `scripts/generate_region_analysis_real_figures.py`. Their input is the tracked CC0 `poly-haven-studio-erp.jpg` derivative, SHA-256 `9b6e2b7521e2cf35d998f4087840c4bf98b480467bd753d719533228c984be8a`. The source is opened with `EquirectangularImage.open` and resized with the sphere-aware container API. The displayed bright-region candidate uses a sphere-aware Gaussian blur followed by a fixed luminance quantile; it is a deterministic integration fixture, **not semantic ground truth**. Watershed regions are geometric partitions and are not presented as object classes. `region-analysis-real-metadata.json` records the exact thresholds, marker strategy, result counts, dimensions and output hashes. Regenerate with: ```bash python scripts/generate_region_analysis_real_figures.py \ --output-dir docs/_static/tutorials ``` ## Application-first multiscene gallery `scripts/generate_application_gallery.py` downloads four pinned 2K HDRIs into an external cache, verifies size and SHA-256 before decoding, applies one fixed tone map, and produces the six tracked application-gallery JPEGs. Use: ```bash python scripts/generate_application_gallery.py \ --cache-dir /path/to/external/poly-haven-cache \ --output-dir docs/_static/tutorials ``` Pass `--offline` to require an already verified cache. Original HDR/EXR files, cache contents, model checkpoints, and intermediate faces are not tracked or distributed. All four assets were retrieved on 2026-10-12 from Poly Haven and are published under [CC0](https://polyhaven.com/license). ### Boiler Room - Asset page: - Author: Oliksiy Yakovlyev - Source file: `boiler_room_2k.hdr` - Source URL: - Source size: 6,707,438 bytes - Source MD5 (upstream API): `324d15c18b121a23fda28e35ba280072` - Source SHA-256: `07c353cb275f1e359cdfa1710acb77c5cd075ce7770859ca56641993263d2d19` The left side of `application-semantics-boiler-room.jpg` is a real 512×1024 spherical ResNet18 run, not a segmentation mock-up. It shows ImageNet class `prison` (index 743), the fixed `grid-like-structure` proxy, and its blue–yellow–red CAM. The class ranked first with probability 0.1920 in this exact run. The dense 16×32 CAM is normalized only within the selected class for display. The right side is an original **method diagram** showing how two registered spherical FOV proposals may support one shared hypothesis; it is not a measured second-view association. The CAM is directional evidence, not ground truth, an object mask, or a calibrated pixel probability. The verified external Torchvision checkpoint is `ResNet18_Weights.IMAGENET1K_V1`, SHA-256 `f37072fd47e89c5e827621c5baffa7500819f7896bbacec160b1a16c560e07ec`. ### Boma - Asset page: - Authors: Greg Zaal (photography), Jarod Guest (processing) - Source file: `boma_2k.hdr` - Source URL: - Source size: 6,954,266 bytes - Source MD5 (upstream API): `385d91640b007c7cacbc334a03549232` - Source SHA-256: `9d2e230d5fdd89efd15b6081a3951c90c771687edd81bc1e08cbb66adb98dd8c` `application-processing-boma.jpg` contains two real 512×1024 sphere-native outputs: constant-angle Gaussian smoothing and geodesic Canny over the source ERP. A 0.60 display gamma makes the dark interior legible; the Canny mask is dilated by three pixels for card-scale display only. The undilated edge count is retained in metadata. ### Mosaic Tunnel - Asset page: - Author: Greg Zaal - Source file: `mosaic_tunnel_2k.hdr` - Source URL: - Source size: 6,625,466 bytes - Source MD5 (upstream API): `a8847cae4e2418f774d5fd6b2cbb4d6b` - Source SHA-256: `47f9267f0620158ccdb7fa7bdf25fd3aa4cd9de3fb88a9bfea68b056c739d6d4` `application-features-mosaic-tunnel.jpg` contains real Stable-core SIFT + FLANN results between the 512×1024 source and a documented 119.8828° cyclic longitude shift. The fixed run retained 262 valid matches and plots the 90 best descriptor distances; ten plotted matches cross the ERP seam under the known rotation. ### Sundowner Overlook - Asset page: - Author: Dario Barresi - Source file: `sundowner_overlook_2k.hdr` - Source URL: - Source size: 6,872,718 bytes - Source MD5 (upstream API): `3ad4d3f49c8751272678a508a2ba77d4` - Source SHA-256: `3cdcb8a8f066e02d5554ce2d6a2591c9a47e02959297f4b2649fd91a49014498` `application-projection-sundowner.jpg` contains the real 512×1024 source with six 95° cube-face footprints, the six 256×256 projected faces, Gaussian-blended ERP reconstruction, round-trip error, and explicit support. The mean supported RGB round-trip error is 2.9485 on the uint8 scale; support covers the full ERP. `pinhole-workflow-sundowner.jpg` is a second real run over the same verified source. From left to right it contains the source ERP, twelve 160×160 icosahedron-sampled gnomonic views at 82° FOV, those views after OpenCV `GaussianBlur` with a 15×15 kernel and sigma 3.0, and the Gaussian-blended ERP reconstruction. The final support fraction is 1.0. Intermediate faces are generated transiently and are not distributed separately. `application-pose-range.jpg` combines an original PanorAi method diagram of two registered spherical cameras and their bearing geometry with the estimated radial-range and relative-error panels from the existing analytic Poly Haven Studio fixture. The diagram is not presented as measured reconstruction data. ### Tracked derivative checksums - `application-projection-sundowner.jpg` (1200×600, 130,026 bytes): `51f48a1030e6fc9a15b7707ce918d487c2ab9d0fd1f422f3dec965ab3152d6b3` - `application-processing-boma.jpg` (1200×600, 67,662 bytes): `b79152afec7776b58cd04c2f9b616c7113333078eb47429c7389b7dc117675b9` - `application-features-mosaic-tunnel.jpg` (1200×600, 148,113 bytes): `54132e985d473013bfdf03a7767e4ca3aadc6981f83039091886e6a143642f2e` - `application-pose-range.jpg` (1200×600, 95,842 bytes): `7a491dc660b359df2f47611411113149aaafeebab91aa652058692909f8803e9` - `application-semantics-boiler-room.jpg` (1200×600, 58,605 bytes): `25c949dda66f0a9043ab6cbb80a72308eadd4a8cea82718fee17d6b62ab4b5be` - `pinhole-workflow-sundowner.jpg` (1200×300, 40,542 bytes): `dee93ce269ba3f95717cd50a31078f9cc0b7a32bedb38b3b15ac0b75b2834a4d` - `application-strategy-routes.svg` (1600×900 view box, 7,274 bytes): `4b6208518d276568946c80d6f0481a93ba1ceb97dca011ca2e24f622c78567ff` - `application-gallery-metadata.json` (17,924 bytes): `550d3a92d9cf3cbdac841870e8da158c1efc81ec0ce25ce4560db3d01ee14d88` The JSON metadata records the per-application working resolution, exact model and checkpoint provenance, processing parameters, counts, output dimensions, and hashes. JPEGs are saved without EXIF metadata. The 16:9 strategy diagram is geometry-led rather than a collage: its white, scientific-figure layout distinguishes projection-domain, sphere-native, and hybrid/multiview execution at normal README scale. ## Nature Reserve Forest `nature-reserve-forest-erp.jpg` and the five derived feature/image-processing figures in this directory come from **Nature Reserve Forest**, photographed by Dimitrios Savva and processed by Jarod Guest. - Asset page: - License: - Source file: `nature_reserve_forest_1k.hdr` - Source URL: - Source size: 1,900,770 bytes - Source MD5: `93c0154883ffed7ff5300692fc7d959d` - Source SHA-256: `6c943ddd683de2f3d9aaa62596961dfccdc9cf206adebfc198e70235ae5707cd` - Retrieved: 2026-10-04 `scripts/generate_documentation_figures.py` verifies the SHA-256 checksum, applies a fixed tone map, and generates the tracked JPEGs. `scripts/generate_core_workflow_figures.py` uses the tracked CC0 derivative as input and produces six additional real outputs: `projected-canny.jpg` (six 256² cube views, 100° FOV, OpenCV Gaussian/Canny and closest fusion), `spherical-gradient-components.jpg` (sphere-native Scharr east, north and magnitude), and the larger 2×2 `spherical-dog-pyramid.jpg`. Their SHA-256 digests are respectively `6f91d7ff0f61672c9769a8ef5ecb03cb8428e9fb88890055e3458c85eb4b9b68`, `ef6a2eb8472e2e3bacf54b970db74e78bb0a5c83a7a0fa3072ad34b085838ef3`, and `41b430745479803c6285071b72a418b7f62012a1fa49c9ed31cbf29fd0391490`. The typed return-path panels are `return-path-points.jpg` (OpenCV GFTT lifted to bearings and reduced by 0.4° angular NMS), `return-path-regions.jpg` (connected edge components lifted as boxes and reduced by solid-angle IoU), and `return-path-view-evidence.jpg` (per-face mean-RGB scores reduced with `coverage_mean`). Their SHA-256 digests are respectively `2637b1269a1e8b4ae18caeb46965509bc7985807de1b6ce3fcdf0adfc17e6b90`, `6edbe44e27122b94367e1236c0644e64a57a26ff3f7bbf47c66acce5e3f345be`, and `fa026637cf008456abe0684ee51a097a5236f8a87bba6518e8b7611bcfa3df85`. ## External-model documentation outputs `scripts/generate_deep_learning_figures.py` verifies the same CC0 input and requires caller-controlled, checksum-verified external caches. It produced: - `depth-anything-views.jpg`: real DA3Metric-Large inference over six 280², 100° tangent views followed by radial-range conversion and Gaussian fusion; SHA-256 `1349cb77a45128ce8b3c7a9c5c91f075e6ecf4479b144dd3e598b3bb773a83ce`. The pinned source and model card declare Apache-2.0. - `metric3d-execution-routes.svg`: original PanorAi diagram of the direct sphere-native and six-face/fusion routes. Its displayed times come from five measured ConvNeXt-Large calls after warm-up on the CC0 Studio ERP. The 203,243,090-parameter model ports 63 convolutions directly and replaces three learned stride-two upsamplers with DC-normalized spherical resize-convolution; no planar spatial layer remains. The Metric3D source is BSD-2-Clause, but the checkpoint repository published no separate model-card license. PanorAi does not redistribute the checkpoint or represent its terms as BSD-2-Clause. The diagram's pinned SHA-256 is recorded in the documentation tests. - `metric3d-large-spherical-1024x2048.jpg`: a PanorAi-generated visualization of a real sphere-native Metric3D Large inference. The input is the CC0 Poly Haven Studio 2K HDRI, 6,602,301 bytes, retrieved 2026-10-14 from , SHA-256 `65ffbd4ffdce0914d61e5c28e078a14491a973cdbebbd82fc3c47b30153955b5`. The source is tone-mapped at native `1024×2048`; the network output has the same shape. The JPEG overlays 82% Turbo-colored radial range and 18% source RGB without spatial resizing. Its SHA-256 is `5770255ad4c072f951a2b7201b407b4e0000ce5642da3f819af7f228a6868b64`. It contains no checkpoint bytes and makes no accuracy or calibration claim. Exact source/checkpoint hashes, timings, route parameters and limitations are recorded in `deep-learning-figure-metadata.json`. Model source, checkpoints, caches, face rasters and raw prediction arrays remain external. The documentation source distribution admits the generated DA3 and Metric3D visualizations plus the original PanorAi diagram. `spherical-image-processing.jpg` contains real spherical Gaussian, Scharr, and Canny outputs. `spherical-histogram-equalization.jpg` contains the real solid-angle equalization, pixel histograms, and latitude weights. `feature-detectors.jpg` shows real PanorAi/OpenCV extractions. `feature-matches.jpg` compares the panorama with a documented cyclic longitude shift of the same CC0 image so readers can see matching across an ERP seam without claiming an independent real capture. `spherical-dog-sift.jpg` shows real direct spherical DoG detections; its descriptors are computed by OpenCV SIFT on one tangent patch per keypoint. `spherical-dog-pyramid.jpg` shows the checksum-pinned tone-mapped ERP followed by three real direct spherical DoG octaves. It is generated by `scripts/generate_core_workflow_figures.py`; each response is displayed as its absolute value normalized at that octave's 99.5th percentile. The derived JPEG SHA-256 is `41b430745479803c6285071b72a418b7f62012a1fa49c9ed31cbf29fd0391490`. The three spherical semantic overlays also derive exclusively from the same tracked CC0 ERP: - `spherical-cam-imagenet-lakeside.jpg`: ImageNet-1K ResNet18 class `lakeside` (index 975, rank 3); - `spherical-cam-places365-forest.jpg`: Places365 ResNet18 scene `forest/broadleaf` (index 150, rank 3); - `spherical-cam-openclip-path.jpg`: OpenCLIP RN50 prompt `a photo of a path` (rank 2 among six declared prompts). All runs used the source `512×1024` ERP and produced native `16×32` dense lattices. The experiment runner min-max-normalized and interpolated each selected channel for display; ImageMagick then converted the generated PNG to metadata-free JPEG at quality 88 without changing its dimensions. No model checkpoint bytes are present in these images. Their SHA-256 checksums are: - `dd0827f6e6cd5301d5ffe14d086cded0ff7ce02e8955d7a95df9b7cd5c65f43c`; - `bcfb1df74b4f57cfacab38d8695015f4eb3ed70d16734b2afe8311ff5d9e9861`; - `baef726a3330669c2d84e711efb8b96e714890d44771f38a14e6d643eff241df`. These visualizations are qualitative evidence only, not semantic masks or pixel probabilities. The model checkpoints remain external under their upstream terms. ## Poly Haven Studio `poly-haven-studio-erp.jpg` and `spherical-stereo-synthetic.png` derive from **Poly Haven Studio**, photographed by Greg Zaal and published under CC0. - Asset page: - License: - Source file: `poly_haven_studio_1k.hdr` - Source URL: - Source size: 1,682,195 bytes - Source MD5: `a1065f613cb6e0388d82a99dcee23d3b` - Source SHA-256: `dfc8505761018d644997a803f332339328af2b53e82023fb3bec37cefdf43b83` - Retrieved: 2026-10-04 `scripts/generate_spherical_stereo_docs_assets.py` verifies the source SHA-256 and applies a fixed tone map. The real panorama is mapped onto an analytic sphere; a deterministic known camera displacement creates the second view and exact radial-range reference. The panel illustrates the stereo implementation without claiming that the analytic geometry was captured in the real office.