2bidoubi/SeeClear-396k

Dataset Structure

SeeClear-396K is a large-scale synthetic dataset accompanying our paper on transparent object depth estimation. It contains paired transparent and opaque renderings generated from identical scene geometry, camera poses, lighting conditions, and object configurations. The only difference between each pair is the material assigned to the target object, enabling explicit supervision for learning transparency-aware representations. For additional details about the dataset and our method, please visit our Project Page and read our Paper.

Dataset Statistics

• 11 transparent object categories
• 10 indoor scenes
• 110 object–scene combinations
• 600 rendering configurations per object–scene combination
• 396,000 rendered images in total

Each rendering configuration contains:

• Transparent RGB image
• Opaque RGB image
• Absolute depth (.exr)
• Relative depth (.png)
• Surface normals (.png)
• Object mask (.png)

---

Scene Overview

The dataset contains ten indoor scenes used for rendering.

Scene	Preview
Scene 1
Scene 2
Scene 3
Scene 4
Scene 5
Scene 6
Scene 7
Scene 8
Scene 9
Scene 10

Object Categories

The dataset contains 11 transparent object categories used throughout the rendered scenes.

Flask	Bottle	Cup	Goblet
Canister	Water Bottle	Plate	Teapot
Glasses	Jar	Vase

---

Archive Organization

The dataset is distributed as 113 ZIP archives.

Standard Archives (110)

Each archive corresponds to one object placed in one scene.

The archive naming convention is

scene_<scene_number>_object_<object_name>.zip

For example,

scene_1_object_bottle.zip
scene_4_object_vase.zip
scene_9_object_glasses.zip

Additional Archives (3)

Three additional archives cover special rendering scenarios:

• Multi-object rendering for Scene 1
• Object-inside-cabinet rendering for Scene 9
• Object-inside-cabinet rendering for Scene 10

These archives follow the same directory organization as the standard archives.

---

Contents of Each Archive

Each archive has the following structure:

scene_<scene_number>_object_<object_name>.zip
├── cam_pose.txt
├── light_pose.txt
├── scale_factors.txt
└── images/

cam_pose.txt

Contains the Blender camera parameters used to render every viewpoint in the archive.

light_pose.txt

Contains the light positions used for all rendered images in the archive.

scale_factors.txt

Lists the anisotropic scaling factors applied during rendering.

Objects are stretched along one of six deformation modes

• x
• y
• z
• xy
• yz
• xz

using five different scaling magnitudes.

---

Image Naming Convention

All rendered images are stored inside the images/ directory.

Filenames follow the convention

<object_state>_<pose_number>_<stretch_dimension>_<stretch_factor>_<light_position>_<output_type>_<optional_blender_suffix>

where

Field	Description
object_state	transparent or opaque
pose_number	Camera viewpoint identifier (e.g. pose03)
stretch_dimension	Stretch direction (x, y, z, xy, yz, xz)
stretch_factor	Applied scaling factor
light_position	Lighting configuration (e.g. lightL, lightR)
output_type	Output modality (mask, normal, relative_depth, absolute_depth, etc.)
optional_blender_suffix	Blender-generated postfix (may not be present)

Example filenames:

opaque_pose03_yz_0.688_lightL_mask_0002.png

transparent_pose02_xz_0.776_lightR.png

The final Blender-generated suffix is optional and should generally be ignored when parsing filenames. It is automatically generated by Blender and does not encode rendering parameters.

---

Rendering Configurations

Each object–scene pair contains 600 rendering configurations, generated by deterministic combinations of

• 10 camera viewpoints
• 2 lighting setups
• 6 anisotropic deformation directions
• 5 deformation magnitudes

The rendering pipeline is fully deterministic, allowing the same configuration to be reproduced when using the recorded rendering parameters.

---

Paired Transparent and Opaque Images

Every transparent rendering has a corresponding opaque rendering with identical

• object geometry
• object deformation
• camera pose
• illumination
• scene layout

Only the object material differs.

This one-to-one correspondence enables supervised learning methods that explicitly model the appearance changes introduced by transparency while keeping all geometric factors fixed.

---

license: mit