rohan9977/segformer-b4-offroad-segmentation

SegFormer-B4 — Offroad Semantic Segmentation

Fine-tuned SegFormer-B4 for semantic segmentation of offroad desert scenes, trained on synthetic data from Duality AI's Falcon simulation platform.

Model Details

• Architecture: SegFormer-B4 (MiT-B4 backbone)
• Parameters: 64M
• Pretrained on: ADE20K (150 classes)
• Fine-tuned on: Duality AI Falcon synthetic desert images
• Input size: 512×512
• Classes: 10

Performance

Metric	Score
Val mIoU	0.5935
Baseline mIoU	0.2478
Improvement	+0.3457
Val Pixel Accuracy	86.17%

Classes

ID	Class	Color
0	Background	Black
1	Trees	Dark Green
2	Lush Bushes	Bright Green
3	Dry Grass	Tan
4	Dry Bushes	Brown
5	Ground Clutter	Olive
6	Logs	Dark Brown
7	Rocks	Gray
8	Landscape	Sandy Brown
9	Sky	Light Blue

Training Details

• Loss: Combined Dice + Focal Loss
• Optimizer: AdamW (lr=3e-5, wd=0.01)
• Scheduler: CosineAnnealingLR
• Augmentation: HorizontalFlip, ColorJitter, CLAHE, HueSaturationValue, RandomFog
• Epochs: 10 (Phase 1) + 10 (Phase 2)
• Hardware: NVIDIA RTX 4050 6GB

Usage

from transformers import SegformerForSemanticSegmentation, SegformerImageProcessor
from PIL import Image
import torch

processor = SegformerImageProcessor.from_pretrained("rohan9977/segformer-b4-offroad-segmentation")
model = SegformerForSemanticSegmentation.from_pretrained("rohan9977/segformer-b4-offroad-segmentation")

image = Image.open("your_image.jpg")
inputs = processor(images=image, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    predicted_mask = outputs.logits.argmax(dim=1)