The Sleeve and Neck Recognition Paper: Spatial-Aware Non-Local Attention for Fashion Landmark Detection by Li et al.

 Please go for this link

🔍 Problem Addressed:

Fashion landmark detection is challenging due to large variation in clothing appearances and non-rigid deformations. Existing deep learning approaches struggle to effectively capture global context and focus on relevant clothing areas.

---

🧠 Proposed Solution:

The authors propose a new module called the Spatial-Aware Non-Local (SANL) block:

• Built on the non-local neural network framework, which captures global dependencies.

• Enhances spatial awareness by incorporating attention maps generated using Grad-CAM from a pretrained classifier.

• Helps the model focus on key garment regions (e.g., sleeves, hems) to better detect landmarks.

---

🏗️ Model Architecture:

1. Backbone: Feature Pyramid Network (FPN) with ResNet-101.

2. SANL blocks: Inserted at four feature levels (strides 4, 8, 16, 32).

3. Attention maps: Generated using Grad-CAM on a ResNet-18 classifier pretrained on DeepFashion categories.

4. Coarse-to-Fine Design:

   • CoarseNet: Learns from broader (larger-σ) heatmaps.

   • FineNet: Refines predictions using narrower (smaller-σ) heatmaps.

---

🧪 Experiments:

• Datasets: DeepFashion-C and FLD (Fashion Landmark Dataset)

• Metric: Normalized Error (NE) – lower is better.

• The proposed method achieves state-of-the-art performance with NE:

  • 0.0299 (DeepFashion-C)

  • 0.0396 (FLD)

• Outperforms previous methods like FashionNet, DFA, DLAN, and AFGN.

---

🔬 Ablation Studies:

• SANL blocks improve performance over base and non-local models.

• Using spatial attention from corresponding feature strides yields better accuracy than using only high-level maps.

• Training only on visible landmarks and using SANL and Coarse-to-Fine together gives the best results.

---

🔄 Generalization:

• SANL blocks also improved performance on fine-grained classification tasks:

  • CUB-200 (Birds) and FoodNet (Chinese dishes)

  • Gains in Top-1 and Top-3 accuracy over baseline and non-local versions.

---

✅ Conclusion:

• SANL blocks effectively blend spatial attention and global context.

• They improve landmark detection and generalize well to other spatially-sensitive tasks.

• No extra annotation needed—only Grad-CAM-based attention maps are used.