TL;DR
PDDM (Pseudo Depth Diffusion Model) eliminates the need for expensive RGB-D sensors by using "pseudo depth" estimated directly from RGB images for semantic segmentation. By aggregating multiple depth estimations and injecting them into a pre-trained diffusion model as "structured noise," the authors achieve new state-of-the-art results on major indoor scene benchmarks.
What Problem Does This Paper Address?
Semantic segmentation in complex indoor scenes is challenging because objects often overlap or share similar colors. While adding a depth (D) channel to standard RGB data helps significantly, it comes with three major headaches:
- Cost: RGB-D cameras are more expensive and less common than standard cameras.
- Alignment: Aligning depth maps with RGB images is difficult due to different sensor positions.
- Data Quality: Raw depth data is often "noisy," containing holes (missing data) or rough edges caused by reflections and shadows.
The authors ask: Can we get the benefits of depth without the hardware? They propose using Pseudo Depth (PD)—depth maps generated by software algorithms from a single RGB image—to replace physical sensors.
Methods
The researchers introduce two primary innovations to make "Pseudo Depth" work effectively:
1. Pseudo Depth Aggregation Module (PDAM)
Not all depth estimation models are created equal. Some excel at edges, while others are better at global geometry. PDAM takes multiple pseudo-depth maps (generated by different "meta" models) and fuses them. It uses spatial and channel attention to weigh the most reliable information from each source, creating a single, high-quality aggregated depth map that is often cleaner and smoother than real sensor data.
2. Pseudo Depth Diffusion Model (PDDM)
The core architecture leverages a large-scale pre-trained Stable Diffusion UNet as a feature extractor. Typically, diffusion models work by removing Gaussian noise from an image.
- The "Aha!" Moment: Instead of adding random Gaussian noise to the RGB features, the authors treat the Pseudo Depth features as "structured noise."
- By injecting depth information into the diffusion process this way, the model naturally learns to "denoise" the scene by using geometric clues, effectively fusing the two modalities without the massive computational overhead of a traditional dual-branch network.
Datasets
The model was evaluated on two gold-standard datasets for indoor scene understanding:
- NYUv2: 1,449 indoor images across 40 semantic classes.
- SUNRGB-D: A larger-scale dataset with over 10,000 images and 37 classes.
- Meta-Depth Sources: To generate the pseudo depth, they used three leading models: Depth Anything, MetaPrompt-SD, and Marigold.
Key Insights and Ideas
- Pseudo Depth > Real Depth: Surprisingly, the authors found that pseudo depth can sometimes outperform real depth. Because pseudo depth is generated by models trained on massive datasets, it is often smoother and lacks the "holes" and sensor noise found in raw hardware scans.
- Diversity is Strength: Aggregating three different depth models yielded better results than using any single one. This suggests that different depth estimation architectures provide complementary geometric "views" of a scene.
- Efficiency through Diffusion: By treating depth as noise in a pre-trained diffusion backbone, they added only 0.37M extra parameters and negligible FLOPs (0.02G) compared to the baseline, making it much more efficient than previous dual-branch fusion methods.
Key Results
PDDM sets a new bar for indoor semantic segmentation:
- NYUv2: Achieved 64.18 mIoU, outperforming the previous state-of-the-art (ODISE) by a significant margin.
- SUNRGB-D: Achieved 55.11 mIoU, a +2.11 improvement over existing methods.
- Ablation Success: The tests showed that even a simple version of their model (using only one pseudo-depth source) still outperformed most traditional RGB-D models that use real sensor data.
Limitations and Future Work
- Inference Speed: While the authors optimized the fusion strategy, the underlying diffusion backbone is still computationally heavier than lightweight real-time networks (like SegFormer).
- Dependency on Meta-Models: The performance of PDDM is inherently tied to the quality of the monocular depth estimation models used. As those models improve, PDDM will likely improve, but it remains a "downstream" dependency.
- Future Directions: The authors suggest exploring how this "depth-as-noise" strategy might work in outdoor autonomous driving scenarios or for other dense prediction tasks like instance segmentation.
Summary updated