Skip to content

Efficient RGB-D Scene Understanding via Multi-task Adaptive Learning and Cross-dimensional Feature Guidance

Conference: CVPR 2025
arXiv: 2603.07570
Code: To be confirmed
Area: Image Segmentation / RGB-D Scene Understanding / Multi-Task Learning
Keywords: RGB-D Scene Understanding, Multi-Task Adaptive Learning, Cross-Dimensional Feature Guidance, Panoptic Segmentation, Efficient Fusion Encoder

TL;DR

An efficient RGB-D multi-task scene understanding network is proposed. It accelerates inference by utilizing redundant features in an improved fusion encoder, introduces a Normalized Focus Channel Layer (NFCL) and Context Feature Interaction Layer (CFIL) for cross-dimensional feature guidance, and designs a multi-task adaptive loss function to dynamically adjust task weights, achieving SOTA performance on NYUv2/SUN RGB-D/Cityscapes.

Background & Motivation

1. Background

Scene understanding, which enables robots to accurately perceive environments, identify objects, and classify scenes, is the foundation of autonomous robotic decision-making. Multi-task learning achieves mutual reinforcement and collaborative optimization among multiple tasks through information sharing and learning mechanisms.

2. Limitations of Prior Work

  • Dual encoder efficiency issue: Seichter et al. extract RGB and depth features separately using a dual encoder, but fail to fully integrate complementary information.
  • Transformer encoder speed issue: Fischedick et al. jointly extract RGB-D information using Swin Transformer v2, but suffer from high matrix calculation and memory access overheads, leading to slow inference.
  • MLP decoder limitation: The MLP decoder has a simple structure, but error information from shallow features can mislead it, and it primarily focuses on global features while ignoring local information.
  • Fixed loss weight issue: Different tasks have significantly different learning difficulties and data distributions. Fixed weights cannot adapt to the dynamically changing task relationships during training.

3. Key Challenge

How to efficiently fuse RGB-D information while ensuring speed, how to guide cross-dimensional features to integrate local and global information, and how to adaptively adjust multi-task learning priorities.

4. Mechanism

Three innovations: (1) An efficient fusion encoder utilizing channel redundancy; (2) Cross-dimensional feature guidance via NFCL and CFIL; (3) A multi-task adaptive loss based on historical performance.

5. Prior Attempts and Limitations

  • Lin et al. used random loss and weighting to avoid human bias, but introduced performance instability.
  • Liu et al. calculated task weights based on training loss to improve stability, but only adjusted weights in the first batch of each iteration, lacking real-time adaptability.
  • Bottleneck modules lead to information loss and decreased feature diversity due to dimensionality reduction.

6. Solution Overview

A unified multi-task network is designed to handle five tasks: semantic segmentation, instance segmentation, orientation estimation, panoptic segmentation, and scene classification. It is collaboratively optimized through the fusion encoder, NFCL, CFIL, and adaptive loss.

Method

Overall Architecture

RGB-D input \(\rightarrow\) improved fusion encoder (4 stages, based on FasterNet-M) \(\rightarrow\) three-branch output: scene classification head (fully connected layer), semantic decoder (MLP + NFCL + CFIL), instance decoder (Non-bottleneck 1D three-layer structure). Semantic segmentation provides foreground masks to instance segmentation, and both are combined to achieve panoptic segmentation. A multi-task adaptive loss is used during training.

Key Design 1: Efficient Fusion Encoder

  • Function: Simultaneously extract complementary features from RGB and depth data to improve inference speed.
  • Mechanism: Exploit the high similarity (redundancy) of feature channels, performing convolutions on only 1/4 of the channels, then concatenating them with the remaining channels, reducing FLOPs to 1/16 of standard convolutions.
  • Design Motivation: Features from different channels are highly similar, eliminating the need to perform convolutions on all channels; reducing memory access frequency significantly improves inference speed.
  • Depth Weight Initialization: Sum the ImageNet pre-trained RGB three-channel weights to initialize depth weights as \(D=(R+G+B)/2\), avoiding additional pre-training.
  • 4-Stage Design: Each stage contains 3/4/18/3 fusion blocks, respectively. The number of blocks is increased in later stages since the image size is smaller.

Key Design 2: Normalized Focus Channel Layer (NFCL)

  • Function: Enhance the representational capacity of shallow encoder features in the semantic decoder.
  • Mechanism: Utilize the absolute value of the scaling factor \(\gamma\) learned in BatchNorm as an indicator of channel importance. After normalization, channel features are re-weighted and re-ordered.
  • Design Motivation: MLP decoders are easily misled by noise/error information in shallow features. NFCL automatically identifies important channels using the BN \(\gamma\) parameter (greater \(\gamma\) = greater variance = contains more key information).
  • Placement: Applied to the first 3 layers of skip connections in the semantic decoder (stage 4 encoder features are already sufficient and require no extra guidance).

Key Design 3: Context Feature Interaction Layer (CFIL)

  • Function: Compensate for the inadequacy of the MLP semantic decoder in fusing local and global information.
  • Mechanism: Capture multi-scale context information using multi-scale adaptive average pooling (\(1\times1\) and \(5\times5\)), compress channels to \(C/2\), perform bilinear upsampling to unify resolution, and concatenate with original features for fusion.
  • Design Motivation: MLP decoders excel at non-linear mapping but primarily focus on global features. CFIL integrates features of different resolutions through multi-scale pooling, improving the ability to distinguish fine structures and boundaries.

Key Design 4: Non-bottleneck 1D Instance Decoder

  • Function: Feature extraction for instance segmentation and orientation estimation.
  • Mechanism: Decompose a \(3\times3\) 2D convolution into two 1D convolutions (\(3\times1 + 1\times3\)) with non-linear activation functions inserted in between.
  • Design Motivation: 3D decomposition reduces parameters by 30% (when kernel size is 3) while increasing non-linear capacity.

Key Design 5: Multi-Task Adaptive Loss

  • Function: Dynamically adjust the loss weights of each task.
  • Mechanism: Calculate the relative loss \(RL_k = L_k/\sum L_t\) for each task after each batch, maintain a historical running average of relative loss \(AvgRL_k\), and update the weights as \(W_k = \max(\bar{W}_k \times (AvgRL_k)^\alpha, W_{\min})\) using an adjustment factor \(\alpha\).
  • Design Motivation: The learning difficulty of different tasks changes dynamically, and fixed weights cannot adapt. Using \(\alpha=0.01\) achieves fine-grained adjustment, and \(W_{\min}=0.1\) prevents any task from being completely ignored.

Loss & Training

  • Semantic segmentation: Cross-entropy loss
  • Instance center: MSE loss
  • Instance offset: MAE loss
  • Orientation estimation: Continuous probability distribution loss based on cos/sin vectors
  • Scene classification: Cross-entropy loss
  • Panoptic segmentation: No additional loss is computed; evaluated during validation

Key Experimental Results

Encoder Comparison (Table 1, NYUv2)

Encoder Instance PQ↑ MAAE↓ Semantic mIoU↑ Inference Speed
Swin v2 58.49 21.09 49.76 Slow
ConvNeXt v2 41.04 31.24 27.69 Medium
MPViT 57.77 21.18 47.44 Relatively Slow
MetaFormer 53.31 23.69 43.27 Medium
Ours 58.59 18.67 46.83 Fast

Model Complexity Comparison (Table 8)

Method Params FLOPs FPS VRAM Sem. mIoU Inst. PQ
EMSAFormer 72.08M 50.66G 16.32 3188M 49.76 58.49
Ours 71.82M 75.28G 20.33 3293M 49.82 59.90

Semantic Segmentation SOTA Comparison

Dataset Method Backbone mIoU↑
NYUv2 EMSAFormer Swin v2 49.76
NYUv2 Ours FasterNet-M 49.82
SUN RGB-D EMSAFormer Swin v2 44.13
SUN RGB-D Ours FasterNet-M 45.56
Cityscapes EMSAFormer Swin v2 60.76
Cityscapes Ours FasterNet-M 65.11

Ablation Study (Table 7, Stepwise Addition of Components to Framework)

Config Inst. PQ↑ Pan. mIoU↑ Sem. mIoU↑ bAcc↑
Baseline (Swin v2) 58.49 50.51 49.76 77.11
+ Fusion Encoder 58.59 47.37 46.83 74.67
+ Adaptive Loss 59.37 48.39 47.72 76.23
+ CFIL 59.25 50.16 49.72 77.00
+ NFCL 59.90 50.21 49.82 76.57

Key Findings

  1. Fusion Encoder is 24.6% faster than Swin v2: FPS increases from 16.32 to 20.33 with fewer parameters (71.82M vs 72.08M).
  2. CFIL achieves significant effects: Semantic mIoU improves by 2.0 percentage points (from 47.72% to 49.72%), outperforming context modules like ASPP, SPPELAN, RFB, etc.
  3. Applying NFCL to the first 3 layers is optimal: Stage 4 encoder features are already sufficient and require no extra guidance.
  4. Multi-task adaptive loss outperforms fixed weights: The model is most balanced when the adjustment factor \(\alpha=0.01\). The adaptive loss model converges faster and more stably.
  5. Non-bottleneck 1D outperforms other extraction modules: Achieves an Instance PQ of 59.25%, outperforming BasicBlock, Bottleneck, MobileBottleneck, and GhostBottleneck.
  6. Indoor-to-outdoor generalization: Achieves an mIoU of 65.11% on Cityscapes, surpassing Lovász (63.06%) and EMSAFormer (60.76%).

Highlights & Insights

  1. Exploiting channel redundancy is simple yet highly efficient: performing convolutions on only 1/4 of the channels reduces FLOPs by 16 times, significantly accelerating inference.
  2. Using BN \(\gamma\) as channel importance is an elegant design: it introduces no extra parameters and reuses the learning results of existing BN layers.
  3. Adaptive loss is based on historical performance rather than a single batch, making it more stable than Liu et al.'s approach.
  4. Unified multi-task framework handles 5 tasks simultaneously (semantic segmentation, instance segmentation, orientation estimation, panoptic segmentation, and scene classification), offering high practicality.
  5. The depth weight initialization \(D=(R+G+B)/2\) is simple but effectively leverages ImageNet pre-training.

Limitations & Future Work

  1. Balance between accuracy and speed: The fusion encoder only samples 1/4 of the channels, which might lose information. Automatically choosing the optimal ratio via Neural Architecture Search (NAS) is worth exploring.
  2. Scalability to high resolutions: The current implementation is difficult to handle ultra-high resolution images/videos.
  3. Assumption of high-quality RGB-D input: Noise in consumer-grade depth sensors such as reflection, transparent surfaces, and boundary sparseness are not handled.
  4. Frame-by-frame independent processing: Temporal consistency is not utilized, which may lead to segmentation flickering in video scenes.
  5. Conflicts might exist in the requirements of different tasks on the fusion encoder; the current shared encoder scheme may not be optimal.
  • Relation to EMSAFormer: This paper uses EMSAFormer as the main baseline, replacing its Swin v2 encoder with the faster fusion encoder, and adding NFCL, CFIL, and adaptive loss.
  • Relation to FasterNet: The fusion encoder is based on the partial convolution concept of FasterNet-M, extending it to RGB-D fusion scenarios.
  • The concept of utilizing BN \(\gamma\) in NFCL is inspired by Network Slimming (channel pruning) but applied in a different context (feature enhancement vs. pruning).

Rating

  • Novelty: ⭐⭐⭐ (The design of individual components is reasonable, but the innovation margin is limited; concepts like channel redundancy utilization, BN \(\gamma\) as channel weights, and adaptive loss have been explored previously.)
  • Experimental Thoroughness: ⭐⭐⭐⭐ (Validated on three datasets, with ablation studies covering the encoder, CFIL, NFCL, decoder, and loss; however, comparisons with other multi-task baselines are relatively sparse.)
  • Writing Quality: ⭐⭐⭐⭐ (Complete and clear structure, rigorous formulas, with discussions on ethics and limitations; some parts could be more concise.)
  • Value: ⭐⭐⭐⭐ (Provides a practical multi-task RGB-D scene understanding solution that balances speed and accuracy, comprehensively surpassing the baseline on three datasets.)