🚗 Attn-LaneNet: Attention-Guided Residual Lane Segmentation Network

Attn-LaneNet is a high-speed, lightweight convolutional neural network designed and trained entirely from scratch (zero pretrained backbones, zero transfer learning) for single-class autonomous lane semantic segmentation.

Developed by Jitrakorn Jansang (นายจิตรกร จันทร์สังข์, Student ID: 6710110055) for Assignment-10: Neural Network From Scratch w/ Custom Dataset (10%) in 2569-241-353 AI Ecosystem, instructed by Dr. Rattachai Wongtanawijit.

---

⚡ Quickstart (Run Inference in 3 Lines)

Option 1: Using AttnLaneNet.from_pretrained

import torch
from model import AttnLaneNet

# 1. Download & load weights directly from Hugging Face Hub
model = AttnLaneNet.from_pretrained("kim2548TH/attn-lanenet-lane-segmentation")

# 2. Forward pass with input (B, 3, 36, 64)
dummy_input = torch.randn(1, 3, 36, 64)
with torch.no_grad():
    pred_mask = (torch.sigmoid(model(dummy_input)) > 0.5).float()
print("Prediction shape:", pred_mask.shape)  # torch.Size([1, 1, 36, 64])

Option 2: Zero-Git 1-File Run (Download Architecture & Weights Directly)

import torch
import importlib.util
from huggingface_hub import hf_hub_download

# Download code & weights
py_file = hf_hub_download("kim2548TH/attn-lanenet-lane-segmentation", "model.py")
weights_file = hf_hub_download("kim2548TH/attn-lanenet-lane-segmentation", "attn_lanenet_weights.pt")

# Dynamic import
spec = importlib.util.spec_from_file_location("attn_model", py_file)
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)

# Instantiate & Load
model = mod.AttnLaneNet()
model.load_state_dict(torch.load(weights_file, map_location="cpu"))
model.eval()

# Predict on (B=1, C=3, H=36, W=64)
dummy_input = torch.randn(1, 3, 36, 64)
output_logits = model(dummy_input)
print("Output logits:", output_logits.shape)

---

🏗️ Model Architecture

Attn-LaneNet is built with strict adherence to real-time embedded edge deployment requirements:

graph TD
    In["Input Tensor: 64x36x3 (RGB)"] --> Stem["Stem Conv + ResBlock (32ch, 64x36)"]
    Stem --> Down1["MaxPool (2x) + 2x ResBlock-SE (64ch, 32x18)"]
    Down1 --> Down2["MaxPool (2x) + 2x ResBlock-SE (128ch, 16x9)"]
    Down2 --> Bot["Bottleneck + Dilated Context (d=1,2,4) (256ch, 16x9)"]
    Bot --> Up1["Bilinear Upsample (2x) + Skip Fusion (64ch, 32x18)"]
    Down1 -.->|Skip Connection| Up1
    Up1 --> Up2["Bilinear Upsample (2x) + Skip Fusion (32ch, 64x36)"]
    Stem -.->|Skip Connection| Up2
    Up2 --> Head["Conv 1x1 Segmentation Head (1ch, 64x36)"]
    Head --> Out["Binary Lane Logits / Mask (64x36x1)"]

Architectural Highlights

1. Zero Spatial Truncation Hierarchy: Native aspect ratio $16:9$ ($64 \times 36 \to 32 \times 18 \to 16 \times 9$) with perfect integer divisibility ($36/2=18, 18/2=9$; $64/2=32, 32/2=16$). No asymmetric padding distortion.
2. Squeeze-and-Excitation (SE) Channel Recalibration: Dynamically focuses network capacity on road lane markings while attenuating reservoir water, sky, and roadside trees.
3. Atrous Spatial Pyramid Context Bottleneck: Multi-dilation convolutional branches ($r = 1, 2, 4$) capture wide road perspective without expanding parameter counts.
4. Ultra-Low Memory Footprint: Only 924,497 parameters and 3.53 MB FP32 RAM, allowing high-frequency real-time execution (>270 FPS) on CPU or edge hardware.

---

📊 Quantitative Evaluation

Evaluated across all 350 unseen test frames (35% split) from the PSU Reservoir Lane Dataset:

---

📉 Loss Convergence Curve

Trained for 35 epochs using Hybrid BCE + Soft Dice Loss with AdamW and Cosine Annealing:

---

📷 Qualitative Inference Snapshot

Qualitative 4-panel visual verification on test frame:

Panel Description: (1) Original Raw Image, (2) Ground Truth YOLO Polygon Annotation, (3) Predicted Binary Lane Mask (Upscaled $1280 \times 720$), (4) Qualitative Overlay (Green=Prediction, Red=Ground Truth, Yellow=Accurate Overlap).

---

📦 Checkpoint Files Available

• model.safetensors (3.6 MB) - SafeTensors binary format
• attn_lanenet_weights.pt (3.6 MB) - PyTorch state dict format
• pytorch_model.bin (3.6 MB) - Standard Hugging Face PyTorch binary format
• config.json - Architecture & metadata configuration
• model.py - Complete standalone PyTorch network definition

---

📚 Dataset

The accompanying dataset with 1,000 frames (650 train / 350 test) annotated with YOLO Class 0 polygons is hosted on Hugging Face:
👉 https://huggingface.co/datasets/kim2548TH/psu-reservoir-lane-dataset

---

👤 Author & Academic Citation

• Author: นายจิตรกร จันทร์สังข์ (Jitrakorn Jansang)
• Student ID: 6710110055
• Course: 2569-241-353 AI Ecosystem (Monday Section)
• Instructor: Dr. Rattachai Wongtanawijit (ดร. รัฐชัย วงศ์ธนวิจิต)
• Institution: Prince of Songkla University (PSU)