[Paper Review] Learning Lightweight Lane Detection CNNs by Self Attention Distillation

Yuenan Hou, Zheng Ma+2 more|arXiv (Cornell University)|Aug 2, 2019

Autonomous Vehicle Technology and Safety23 references96 citations

TL;DR

The paper introduces Self Attention Distillation (SAD), a training-time technique that uses a network’s own attention maps as soft targets to improve representation learning, enabling lightweight lane-detection CNNs (e.g., ENet-SAD) to match or surpass heavier models without extra labels or slower inference.

ABSTRACT

Training deep models for lane detection is challenging due to the very subtle and sparse supervisory signals inherent in lane annotations. Without learning from much richer context, these models often fail in challenging scenarios, e.g., severe occlusion, ambiguous lanes, and poor lighting conditions. In this paper, we present a novel knowledge distillation approach, i.e., Self Attention Distillation (SAD), which allows a model to learn from itself and gains substantial improvement without any additional supervision or labels. Specifically, we observe that attention maps extracted from a model trained to a reasonable level would encode rich contextual information. The valuable contextual information can be used as a form of 'free' supervision for further representation learning through performing topdown and layer-wise attention distillation within the network itself. SAD can be easily incorporated in any feedforward convolutional neural networks (CNN) and does not increase the inference time. We validate SAD on three popular lane detection benchmarks (TuSimple, CULane and BDD100K) using lightweight models such as ENet, ResNet-18 and ResNet-34. The lightest model, ENet-SAD, performs comparatively or even surpasses existing algorithms. Notably, ENet-SAD has 20 x fewer parameters and runs 10 x faster compared to the state-of-the-art SCNN, while still achieving compelling performance in all benchmarks. Our code is available at https://github.com/cardwing/Codes-for-Lane-Detection.

Motivation & Objective

• Motivate robust lane detection under sparse supervisory signals and challenging conditions (occlusion, low lighting).
• Develop a self-supervised distillation mechanism (SAD) that leverages internal attention maps for training.
• Demonstrate that SAD improves lightweight architectures without increasing inference time.
• Show that SAD yields state-of-the-art or competitive performance on major lane-detection benchmarks.
• Provide architectural refinements to ENet to maximize SAD benefits.

Proposed method

• Define attention maps from activation A_m at layer m and construct attention via G(A_m) using channel-wise statistics (prefer G_sum^2).
• Introduce an attention generator Ψ that produces distilled targets Ψ(A_m) for layer m+1; use spatial softmax and bilinear upsampling as needed.
• Formulate layer-wise, top-down distillation loss L_distill(A_m, A_{m+1}) = sum_m L_d(Ψ(A_m), Ψ(A_{m+1})) with L_d typically an L2 loss.
• Combine standard segmentation loss (cross-entropy) with IoU loss and a lane-existence loss, plus the distillation loss: L = L_seg + α L_IoU + β L_exist + γ L_distill.
• Apply SAD to ENet-based architectures by inserting an AT-GEN after encoder blocks and explore different distillation paths and timepoints during training.
• Ablate distillation paths and compare SAD with deep supervision to assess effectiveness.

Experimental results

Research questions

• RQ1 Can a model learn useful lane-context representations by distilling its own attention maps across layers (self-distillation) without external labels?
• RQ2 Which layer-wise mimicking paths (which blocks distill to which) maximize gains, and when during training should SAD be introduced?
• RQ3 Does SAD provide benefits over traditional deep supervision and does it maintain fast inference in lightweight networks?
• RQ4 How does SAD affect attention maps and resulting lane-detection performance across diverse benchmarks (TuSimple, CULane, BDD100K)?

Key findings

• SAD consistently improves performance of lightweight networks on three benchmarks (TuSimple, CULane, BDD100K).
• ENet-SAD achieves state-of-the-art-like results among lightweight models, with 20x fewer parameters and 10x faster runtime than SCNN on CULane, while maintaining strong accuracy.
• On TuSimple, ENet-SAD attains 96.64% accuracy (vs. 93.02% for ENet baseline); on CULane, it reaches a higher F1 in the reported comparisons; on BDD100K, ENet-SAD achieves 36.56% IoU and 16.02% IoU for the respective metrics shown.
• Deeper networks also benefit from SAD (e.g., ResNet-101 with SAD improves CULane F1 from 70.8 to 71.8 and BDD100K accuracy from 34.45% to 35.56%).
• Ablation shows middle/high-level layer targets and neighboring-layer mimicking yield the best gains; backward distillation (higher-to-lower) degrades performance.
• SAD provides gains beyond deep supervision, delivering softer, context-rich targets and a top-down feedback mechanism that enhances both shallow and deep representations.