mmseg-baseline

🛰️ mmseg-baseline (DeepLabV3+ R50 on UAV Dataset)

🧠 Model Description

This model is a semantic segmentation model trained with MMSegmentation, based on:

• Architecture: DeepLabV3+
• Backbone: ResNet-50 (dilated, stride-8)
• Framework: MMSegmentation
• Input size: 512 × 512
• Training schedule: 80k iterations

The model is designed for urban scene understanding in UAV imagery, supporting multi-class segmentation.

---

📊 Dataset

• Dataset: UAV segmentation dataset (CamVid-style format)
• Number of classes: 12

Classes

---

🗂️ Dataset Structure

data/uav/
├── images/
│   ├── train/
│   └── val/
├── annotations/
│   ├── train/
│   └── val/
├── train.txt
└── val.txt

---

⚙️ Training Details

• Framework: MMSegmentation

• Optimizer: (default from base config, typically SGD or AdamW)

• Batch size: 2

• Resize scale: (960, 720)

• Augmentations:

  • Random resize (0.5–2.0)
  • Random crop (512×512)
  • Random flip (p=0.5)

---

📈 Evaluation

• Metric: mIoU (Mean Intersection over Union)
• Evaluator: IoUMetric

⚠️ Please fill in actual results if available.

Example:

---

🚀 Usage

1️⃣ Download model

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Thengy/mmseg-baseline",
    local_dir="model"
)

---

2️⃣ Inference with MMSegmentation

from mmseg.apis import init_model, inference_model

config = "config.py"
checkpoint = "checkpoints/iter_80000.pth"

model = init_model(config, checkpoint, device="cuda:0")

result = inference_model(model, "demo.jpg")

---

🧪 Training

To reproduce training:

cd mmsegmentation

python tools/train.py path/to/config.py

---

⚠️ Limitations

• Performance may degrade under:

  • unseen environments
  • different UAV altitudes or camera angles

• Sensitive to annotation noise

• Limited generalization outside training domain

---

📄 License

MIT

---

🙋 Author

• Author: ThengAndrew
• Date: 2026-04-30

---

📌 Notes

• The configuration is adapted from:

  • DeepLabV3+ R50 MMSegmentation baseline

• Ensure consistency between:

  • num_classes
  • dataset annotations
  • label encoding

---

📷 Example Results