MicroDet – Lightweight Drone-Based Object Detection

A lightweight, anchor-free object detection system built using MicroDet, optimized for aerial / drone imagery. This project focuses on efficient person detection with minimal computational overhead, making it suitable for edge devices and UAV applications.

---

Executive Summary

MicroDet delivers efficient, anchor-free person detection for drone imagery with a practical train-validate-infer pipeline. The repository is designed for experimentation speed, stable training behavior, and deployment-friendly inference.

Focus Area	Value
Detection Strategy	Anchor-free regression with DFL
Deployment Goal	Lightweight edge/UAV compatibility
Data Standard	COCO-style annotations
Training Stability	AMP + EMA + configurable assigner/loss

Quick Navigation

• Features
• Project Structure
• Model Overview
• Dataset
• Training
• Evaluation
• Inference
• Configuration Highlights
• Tech Stack
• Applications
• Future Work
• Author - Ravindran S

Quick Start Workflow

1) Install dependencies

python -m pip install -r requirements.txt

2) Configure experiment

Update microdet.toml for dataset paths, optimizer settings, assigner radius, and loss options.

3) Train and validate

Use the training commands in the Training section to start from scratch or resume from a checkpoint.

4) Run inference

Use the inference command in the Inference section to generate visual outputs with NMS-applied detections.

5) Review outputs

Check model checkpoints in runs/microdet_drone/weights/ and logs in runs/microdet_drone/logs/.

---

Features

• Lightweight MicroDet architecture

• Anchor-free detection (DFL-based regression)

• COCO-format dataset support

• Mixed-precision (AMP) training

• EMA (Exponential Moving Average) weights

• End-to-end pipeline: Train → Validate → Infer

• Bounding box visualization with NMS

• Config-driven (.toml) model & training setup

Project Structure

microdet/
├── tmp/
│   ├── model/
│   │   ├── backbone/
│   │   ├── neck/
│   │   ├── detect/
│   │   ├── loss/
│   │   └── model_wrapper.py
│   ├── train/
│   │   ├── train.py
│   │   └── validate.py
│   ├── infer/
│   │   ├── run_infer.py
│   │   └── image.png
│   └── data/
│       ├── coco_dataset.py
│       └── collate.py
├── runs/
│   └── microdet_drone/
│       ├── weights/
│       │   ├── last.ckpt
│       │   └── best.ckpt
│       └── logs/
├── microdet.toml
├── requirements.txt
└── README.md

Model Overview

MicroDet is a one-stage, anchor-free detector designed for speed and efficiency.

Architecture

• Backbone: Lightweight CNN for feature extraction

• Neck: Multi-scale feature aggregation

• Head:

  • Classification branch (Quality Focal Loss)

  • Regression branch (Distribution Focal Loss – DFL)

Feature Map Strides

[8, 16, 32]

Loss Functions

Loss Type	Purpose
Quality Focal Loss (QFL)	Classification confidence
Distribution Focal Loss (DFL)	Bounding box regression
GIoU Loss	Box overlap accuracy

Configured in microdet.toml:

[model.head.loss]
config = [2.0, 0.25, 1.0, 7, "giou"]

Dataset

• COCO-style annotation format

• Single class: person

• Input resolution: 640×640

• Supports training & validation splits

Example:

[data.train]
config = [
  "tmp/data/dataset/images",
  "tmp/data/dataset/result.json",
  [640, 640],
  true,
  {}
]

Training

Train from scratch

python -m src.train.train \
  --config microdet.toml \
  --device cuda

Resume training

python -m src.train.train \
  --config microdet.toml \
  --device cuda \
  --resume runs/microdet_drone/weights/last.ckpt

Evaluation

• Validation runs every val_interval epochs

• Metrics:

  • mAP

  • Confidence stability

  • Qualitative bounding box accuracy

Inference

Run inference on a test image:

python tmp/infer/run_infer.py

Output

• Bounding boxes drawn on original image

• Non-Maximum Suppression (NMS) applied

Output saved to:

tmp/infer/output.png

Sample Output

✔ Detected persons from aerial view ✔ Bounding boxes after NMS ✔ Scaled correctly to original image resolution

Early training may show many low-confidence boxes; tuning assigner radius and confidence threshold improves results.

Configuration Highlights

Assigner (Important)

[model.head.assigner_cfg]
config = ["CenterAssigner", {
  "8"  = 5.0,
  "16" = 5.0,
  "32" = 5.0
}]

Optimizer

[schedule.optimizer]
config = ["adamw", 0.0005, 0.05, true, true]

Known Challenges

• Low confidence during early epochs

• Over-detection without NMS

• DFL decoding tensor contiguity issues

• Correct stride handling during inference

✔ All addressed through architectural tuning and post-processing.

Tech Stack

• Language: Python

• Framework: PyTorch

• Vision: OpenCV

• Model: MicroDet

• Data Format: COCO

• Hardware: CUDA GPU

Applications

• Drone surveillance

• Crowd monitoring

• Search & rescue

• Smart city analytics

• Edge AI deployments

Future Work

• Multi-class detection

• Video inference & tracking

• Model quantization

• Edge deployment (Jetson / TPU)

• Knowledge distillation

Author - Ravindran S

Developer • ML Enthusiast • Neovim Customizer • Linux Power User

Hi! I'm Ravindran S, an engineering student passionate about:

• Linux & System Engineering
• AIML (Artificial Intelligence & Machine Learning)
• Full-stack Web Development
• Hackathon-grade project development

🔗 Connect With Me

You can reach me here:

Socials

Contact