rtarun1/mirage18k

[IROS 2026] Mirage 18k: Dataset for Glass Segmentation & Depth Estimation

Mirage 18k is a novel, multi-task dataset comprising 18,353 manually annotated images across 38 unique indoor scenes, designed specifically for joint glass segmentation and glass-aware monocular depth estimation in robotics.

It contains diverse real-world glass structures (indoor panes, frosted doors, windows, clear doors) with severe background clutter, saliency, and dynamic obstacles.

• Model Checkpoint: SILICA Model Card
• Code Repository: GitHub Repository (rtarun1/Silica)
• Project Page: silica-mirage.github.io
• Example Data: GitHub example/ Directory

This work has been accepted for publication at IROS 2026, as part of the work SILICA: Repurposing Diffusion Priors for Joint Glass Segmentation and Depth Estimation.

---

Dataset Acquisition & Overview

Acquisition Protocol

Ground-truth depth for transparent surfaces is captured by modeling each glass pane as a 3D planar surface. Temporary opaque markers are placed at pane corners, and their median depth is extracted from local neighborhoods. A stationary camera records the scene twice—once with markers to capture accurate depth, and once without markers from the exact same pose to obtain clear RGB imagery. Reconstructions are finalized via least-squares optimization in inverse depth space.

Diversity & Scene Distribution

Mirage 18k covers diverse distributions of transparent glass structures in wild indoor environments, including indoor panes, frosted doors, windows, clear doors, and glass-free negative samples.

---

Repository Structure

The dataset is organized into timestamped sequence folders representing individual capture scenes, accompanied by split definition text files at the root directory:

Mirage18k/
├── seg_train_images.txt
├── seg_train_masks.txt
├── seg_test_images.txt
├── seg_test_masks.txt
├── depth_test_rs_rgb.txt
├── depth_test_binary_mask.txt
├── depth_test_rs_raw_depth_image.txt
├── depth_test_rs_gt_depth_image.txt
└── YYYY-MM-DD_HH-MM-SS/                    # Timestamped scene folders (38 total)
    ├── rs_images/                          # Standard 8-bit RGB images (.png)
    ├── binary_masks/                       # 1-channel binary glass segmentation masks (.png)
    ├── rs_depth_img/                       # 16-bit raw sensor depth images (.png)
    └── rs_corresponding_gt/                # 16-bit planar ground-truth glass depth (.png)

Dataset Splits

To guarantee strict out-of-distribution evaluation, scenes included in test splits do not overlap with training splits.

• Segmentation Train Split: 12,420 samples (seg_train_images.txt, seg_train_masks.txt)

• Segmentation Test Split: 3,527 samples (seg_test_images.txt, seg_test_masks.txt)

• Depth & Joint Evaluation Split: 2,406 samples with paired raw sensor depth and ground-truth depth (depth_test_*.txt)

Quick Test: To run quick standalone inference on a sample subset without downloading the full dataset, check out the sample images, raw depth, and camera intrinsic files available directly in our GitHub example/ folder.