OpenRooms Dataset Release

Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, Sai Bi, Zexiang Xu, Hong-Xing Yu, Kalyan Sunkavalli, Miloš Hašan, Ravi Ramamoorthi, Manmohan Chandraker

News

[09/10/21]. All the rendered ground-truths are available for downloading.

[08/07/21]. Please send an email to [email protected] if you hope to receive the newest update.

[06/29/21]. We released two demo scenes with all ground-truths (Demo and Demo.zip). Please use the two scenes to determine usage for your applications.

[05/19/21]. We released all rendered images (Images and Images.zip).

About

This is the release webpage for the OpenRooms dataset. We first introduce the rendered images and various ground-truths. Next, we introduce how to render your own images based on the OpenRooms dataset creation framework. For each type of data, we offer two kinds of formats, zip files and individual folders, so users may choose whether to download the whole dataset more efficiently or download individual folders for different scenes. We recommend using Rclone to avoid slow or unstable downloads.

OpenRooms is a collaboration between researchers from UCSD and Adobe. We acknowledge generous support from NSF, ONR, Adobe and Google. For any questions, please email: [email protected].

Dataset Overview

We render six versions of images for all the scenes. Those rendered results are saved in 6 folders: main_xml, main_xml1, mainDiffMat_xml, mainDiffMat_xml1, mainDiffLight_xml and mainDiffLight_xml1. All 6 versions are built with the same CAD models. main_xml, mainDiffMat_xml, mainDiffLight_xml share one set of camera views while main_xml1, mainDiffMat_xml1 and mainDiffLight_xml1 share the other set of camera views. main_xml(1) and mainDiffMat_xml(1) have the same lighting but different materials while main_xml(1) and mainDiffLight_xml(1) have the same materials but different lighting. Both the lighting and material configuration of main_xml and main_xml1 are different. We believe this configuration can potentially help us develope novel applications for image editing. Two example scenes from main_xml, mainDiffMat_xml and mainDiffLight_xml are shown in the below.

News: We currently only release the rendered images of the dataset. All ground-truths will be released in a few days. The dataset creation pipeline will also be released soon.

Rendered Images and Ground-truths

All rendered images and the corresponding ground-truths are saved in folder data/rendering/data/. In the following, we will detail each type of rendered data and how to read and interpret them. Two example scenes with images and all ground-truths are included in Demo and Demo.zip. The training/testing split of the scenes can be found in train.txt and test.txt.

1. Image and Image.zip: The 480 × 640 HDR images im_*.hdr, which can be read with the python command.

   im = cv2.imread('im_1.hdr', -1)[:, :, ::-1]

   We render images for main_xml(1), mainDiffMat_xml(1) and mainDiffLight_xml(1).

2. Material and Material.zip: The 480 × 640 diffuse albedo maps imbaseColor_*.png and roughness map imroughness_*.png. Note that the diffuse albedo map is saved in sRGB space. To load it into linear RGB space, we can use the following python commands. The roughness map is saved in linear space and can be read directly.

   im = cv2.imread('imbaseColor_1.hdr')[:, :, ::-1]
   im = (im.astype(np.float32 ) / 255.0) ** (2.2)

   We only render the diffuse albedo maps and roughness maps for main_xml(1) and mainDiffMat_xml(1) because mainDiffLight_xml(1) share the same material maps with the main_xml(1).

3. Geometry and Geometry.zip: The 480 × 640 normal maps imnomral_*.png and depth maps imdepth_*.dat. The R, G, B channel of the normal map corresponds to right, up, backward direction of the image plane. To load the depth map, we can use the following python commands.

   with open('imdepth_1.dat', 'rb') as fIn:
       # Read the height and width of depth
       hBuffer = fIn.read(4)
       height = struct.unpack('i', hBuffer)[0]
       wBuffer = fIn.read(4)
       width = struct.unpack('i', wBuffer)[0]
       # Read depth
       dBuffer = fIn.read(4 * width * height )
       depth = np.array(
           struct.unpack('f' * height * width, dBuffer ),
           dtype=np.float32 )
       depth = depth.reshape(height, width)

   We render normal maps for main_xml(1) and mainDiffMat_xml(1), and depth maps for main_xml(1).

4. Mask and Mask.zip: The 480 × 460 grey scale mask immask_*.png for light sources. The pixel value 0 represents the region of environment maps. The pixel value 0.5 represents the region of lamps. Otherwise, the pixel value will be 1. We render the ground-truth masks for main_xml(1) and mainDiffLight_xml(1).

5. SVLighting and SVLighting.zip: The (120 × 16) × (160 × 32) per-pixel environment maps imenv_*.hdr. The spatial resolution is 120 x 160 while the environment map resolution is 16 x 32. To read the per-pixel environment maps, we can use the following python commands.

   # Read the envmap of resolution 1920 x 5120 x 3 in RGB format
   env = cv2.imread('imenv_1', -1)[:, :, ::-1]
   # Reshape and permute the per-pixel environment maps
   env = env.reshape(120, 16, 160, 32, 3)
   env = env.transpose(0, 2, 1, 3, 4)

   We render per-pixel environment maps for main_xml(1), mainDiffMat_xml(1) and mainDiffLight_xml(1). Since the total size of per-pixel environment maps is 4.0 TB, we do not provide an extra .zip format for downloading. Please consider using the tool Rclone if you hope to download all the per-pixel environment maps.

6. SVSG and SVSG.zip: The ground-truth spatially-varying spherical Gaussian (SG) parameters imsgEnv_*.h5, computed from this optimization code. We generate the ground-truth SG parameters for main_xml(1), mainDiffMat_xml(1) and mainDiffLight_xml(1). For the detailed format, please refer to the optimization code.

7. Shading and Shading.zip: The 120 × 160 diffuse shading imshading_*.hdr computed by intergrating the per-pixel environment maps. We render shading for main_xml(1), mainDiffMat_xml(1) and mainDiffLight_xml(1).

8. SVLightingDirect and SVLightingDirect.zip: The (30 × 16) × (40 × 32) per-pixel environment maps with direct illumination imenvDirect_*.hdr only. The spatial resolution is 30 × 40 while the environment maps resolution is 16 × 32. The direct per-pixel environment maps can be load the same way as the per-pixel environment maps. We only render direct per-pixel environment maps for main_xml(1) and mainDiffLight_xml(1) because the direct illumination of mainDiffMat_xml(1) is the same as main_xml(1).

9. ShadingDirect and ShadingDirect.zip: The 120 × 160 direct shading imshadingDirect_*.rgbe. To load the direct shading, we can use the following python command.

   im = cv2.imread('imshadingDirect_1.rgbe', -1)[:, :, ::-1]

   Again, we only render direct shading for main_xml(1) and mainDiffLight_xml(1)

10. SemanticLabel and SemanticLabel.zip: The 480 × 640 semantic segmentation label imsemLabel_*.npy. We provide semantic labels for 45 classes of commonly seen objects and layout for indoor scenes. The 45 classes can be found in semanticLabels.txt. We only render the semantic labels for main_xml(1).

11. LightSource and LightSource.zip: The light source information, including geometry, shadow and direct shading of each light source. In each scene directory, light_x directory corresponds to im_x.hdr, where x = 0, 1, 2, 3 ... In each light_x directory, you will see files with numbers in their names. The numbers correspond to the light source ID, i.e. if the IDs are from 0 to 4, then there are 5 light sources in this scene.

    • Geometry: We provide geometry annotation for windows and lamps box_*.dat for main_xml(1) only. To read the annotation, we can use the following python commmands.

      with open('box_0.dat', 'rb')  as fIn:
          info = pickle.load(fIn )

      There are 3 items saved in the dictionary, which we list blow.
      • isWindow: True if the light source is a window, false if the light source is a lamp.
      • box3D: The 3D bounding box of the light source, including center center, orientation xAxis, yAxis, zAxis and size xLen, yLen, zLen.
      • box2D: The 2D bounding box of the light source on the image plane x1, y1, x2, y2.
    • Mask: The 120 × 160 2D binary masks for light sources mask*.png. We only provide the masks for main_xml(1).
    • Direct shading: The 120 × 160 direct shading for each light source imDS*.rgbe. We provide the direction shading for main_xml(1) and mainDiffLight_xml(1).
    • Direct shading without occlusion: The 120 × 160 direct shading without occlusion for each light source imNoOcclu*.rgbe. We provide the direction shading for main_xml(1) and mainDiffLight_xml(1).
    • Shadow: The 120 × 160 shadow maps for each light source imShadow*.png. We render the shadow map for main_xml(1) only.

12. Friction and Friction.zip: The friction coefficients computed from our SVBRDF following the method proposed by Zhang et al. We compute the friction coefficients for main_xml(1) and mainDiffLight_xml(1)

Dataset Creation

1. GPU renderer: The Optix-based GPU path tracer for rendering. Please refer to the github repository for detailed instructions.
2. Tileable texture synthesis: The tielable texture synthesis code to make sure that the SVBRDF maps are tileable. Please refer to the github repository for more details.
3. Spherical gaussian optimization: The code to fit per-pixel environment map with spherical Gaussian lobes, using LBFGS optimization. Please refer to the github repository for detailed instructions.

The CAD models, environment maps, materials and code required to recreate the dataset will be released soon.

Applications

1. Inverse Rendering: Trained on our dataset, we achieved state-of-the-arts on some inverse rendering metrics, especially the lighting estimation. Please refer to our github repository for the training and testing code.
2. Robotics: Our robotics applications will come soon.

Related Datasets

The OpenRooms dataset is built on several prior works, as noted below. Please refer to them for the respective license details and terms of use when creating your own dataset using OpenRooms.

1. ScanNet dataset: The real 3D scans of indoor scenes.
2. Scan2cad dataset: The alignment of CAD models to the scanned point clouds.
3. Laval outdoor lighting dataset: HDR outdoor environment maps
4. HDRI Haven lighting dataset: HDR outdoor environment maps
5. PartNet dataset: CAD models
6. Adobe Stock: High-quality microfacet SVBRDF texture maps. Please license the materials from Adobe Stock.