Semantic Distillation Architecture

• Semantic distillation architecture is a framework that integrates proxy semantic supervision with geometric and motion cues to achieve comprehensive scene understanding.
• The design leverages a modular setup with DSNet, CamNet, and OFNet, enabling joint depth, pose, and optical flow estimation through self-distillation and cross-task consistency.
• It delivers efficient, real-time performance in resource-constrained settings, making it ideal for autonomous navigation, AR/VR, and embedded applications.

Semantic distillation architecture encompasses a family of neural network design and training strategies in which semantic or contextual knowledge—often available only indirectly or via proxy—is transferred, distilled, or otherwise embedded into a compact or compositional representation enabling comprehensive multi-task understanding. In the architecture proposed by "Distilled Semantics for Comprehensive Scene Understanding from Videos" (Tosi et al., 2020), this concept is realized through a multi-task system for real-time, end-to-end scene understanding from monocular video, where geometric, motion, and semantic cues are learned jointly. Distillation is achieved by leveraging proxy semantic labels generated by a pre-trained segmentation network, combined with self-supervised geometric learning and a self-distillation protocol for optical flow, resulting in a lightweight model suitable for power- and resource-constrained scenarios.

1. Multi-Task Modular Design

The semantic distillation architecture consists of three interacting subnetworks, each dedicated to a specific modality but structurally coupled for information sharing:

• Depth and Semantic Network (DSNet): A compact encoder-decoder backbone (inspired by PydNet) predicts per-frame depth ($D_t$) and semantic masks ($S_t$) using a pyramidal feature extractor down to 1/32 spatial resolution. Coarse depth is predicted at lowest resolution and refined by upsampling and concatenating features from higher (coarser) levels. At half-resolution, two separate predictors output final depth and semantic segmentation masks. Feature sharing between these two tasks enables joint reasoning and mutual supervision, e.g., semantic object boundaries help depth estimation and depth edges refine semantic segmentation.
• Camera Network (CamNet): Estimates camera pose ($T_{t \to s}$) and intrinsics ($K$) between video frames. It uses a dual encoder design in which features from target and source images are separately extracted and concatenated to regress translation vectors, Euler angles, and intrinsic parameters. This provides geometric context crucial for disentangling background motion from object motion in scenes observed by a moving monocular camera.
• Optical Flow Network (OFNet) with Self-Distillation: A lightweight PWC-Net variant predicts dense 2D optical flow ($F_{t \to s}$) between frames. OFNet’s training is coupled to the outputs of DSNet and CamNet (by considering the rigid flow induced by the scene geometry and relative pose) and incorporates semantic segmentation predictions. A distinct self-distillation regime is employed whereby a refined flow predictor (SD-OFNet) learns to reconcile teacher flow with “grounded” rigid flow, especially to handle occlusion and non-rigid object motion.

This modular composition supports the simultaneous output of depth, semantics, per-pixel optical flow, motion probabilities, and instance-level motion masks at real-time speed.

2. Joint Training Protocol with Proxy Semantic Supervision

Learning proceeds in staged fashion:

• Stage 1: DSNet and CamNet are co-trained unsupervisedly for geometry and semantics. Depth prediction relies on photometric reconstruction: given a target $I_t$ and source $I_s$, the network uses its depth, relative pose, and intrinsics to generate a warped source frame as a reconstruction of $I_t$. The mapping from $p_t$ in $I_t$ to $S_t$0 in $S_t$1 is:

$S_t$2

The photometric loss is:

$S_t$3

where $S_t$4 incorporates terms such as SSIM and the $S_t$5 difference.

• Proxy Semantics: Semantic segmentation supervision is provided via soft targets generated by an offline state-of-the-art segmentation model (e.g., pre-trained on Cityscapes), which “distills” its knowledge to the compact DSNet during training. Cross-entropy loss is computed between DSNet’s predictions and the proxy labels. Additionally, a cross-task edge consistency loss ($S_t$6) forces depth edges to align with semantic boundaries, ensuring that geometric cues inform high-level semantics.
• Stage 2: OFNet is first trained as a teacher with self-supervised optical flow loss, but student flow is often unreliable near occluded or non-rigid regions. To correct this, a rigid flow $S_t$7 is computed from geometry and pose:

$S_t$8

Self-distillation is subsequently applied, where the training loss guides the student flow to align with $S_t$9 in inconsistent regions and with the teacher prediction elsewhere, using a consistency mask $T_{t \to s}$0 based on motion probability and semantic priors. The final flow loss is

$T_{t \to s}$1

where $T_{t \to s}$2 is $T_{t \to s}$3 distance and $T_{t \to s}$4 are task balances.

3. Semantic Distillation via Proxy Labels and Edge Consistency

Semantic cues are distilled by:

• Generating pixelwise semantic proxy labels $T_{t \to s}$5 using a pre-trained segmentation model. These serve as cost-effective “ground truth” for training DSNet.
• Training with a standard cross-entropy loss $T_{t \to s}$6 and enforcing a cross-task depth-edge-to-semantic-boundary alignment with $T_{t \to s}$7.
• The distillation process is soft, using probability maps rather than hard labels, thus providing richer supervisory information for semantic segmentation.
• The propagated semantic information further improves both geometric and optical flow estimation, as learning semantic object extents and categories provides spatial and contextual priors that constrain depth discontinuities and motion segmentation.

4. Mathematical Formulation of Core Computations

The network’s learning framework is grounded in the following key equations:

• Image warping by geometry and pose:

$T_{t \to s}$8

• Photometric reconstruction loss:

$T_{t \to s}$9

• Pixelwise motion probability (distinguishing moving from static regions):

$K$0

where $K$1 quantifies angular similarity of flows and $K$2 is a norm ratio.

• Consistency masking for self-distillation:

$K$3

with $K$4 a threshold and $K$5 composed by semantic and boundary masks.

• Final self-distillation loss:

$K$6

These computations explicitly intertwine geometry, semantics, and flow, supporting end-to-end learning.

5. Benchmarks, Empirical Results, and Resource Requirements

The evaluated architecture on the KITTI and Cityscapes benchmarks demonstrates the following:

• Monocular depth estimation achieves lower Absolute Relative (Abs Rel), Squared Rel, and RMSE errors than competitors at long-range distances, using less than 8.5M parameters.
• Optical flow estimation, measured by the F1 score on KITTI 2015, yields superior results compared to prior self-supervised multi-task methods.
• Motion segmentation performance, derived from motion probabilities and semantic priors, achieves higher pixel accuracy, mean accuracy, and mean Intersection-over-Union (IoU) for moving objects.

The model runs at approximately 60 FPS on Titan Xp GPUs and 5 FPS on embedded Jetson TX2, establishing real-time capabilities even on low-power hardware.

6. Applications, Deployment, and Implications

Semantic distillation architectures of this design are particularly relevant for:

• Autonomous systems: Unified geometric, semantic, and motion understanding enables robust navigation and perception in self-driving and robotic platforms, reducing sensor complexity by leveraging monocular inputs and minimizing annotation demands via proxy supervision.
• Augmented/Virtual Reality (AR/VR): Real-time depth, semantics, and motion outputs support immersive scene interaction and dynamic object rendering in resource-constrained wearable AR/VR hardware.
• Embedded and mobile computing: The architecture’s compactness and low power consumption facilitate deployment on drones, mobile robots, and edge devices.
• Annotation cost savings: Training protocols based on self-supervision and knowledge distillation from offline proxy labels obviate the need for dense ground-truth, accelerating adaptation to novel environments.

In summary, semantic distillation architecture integrates geometric and semantic reasoning through multi-task, proxy-supervised joint learning, delivering real-time and robust scene parsing suitable for embedded and autonomous visual systems.