INR-Enhanced C2f for Small-Object UAV Datasets

README for GitHub. Covers environment setup, datasets, configs, training, and tuning for INR-Enhanced C2f modules (AI-TOD, UAVDT, VisDrone).

1) Environment

# Python >= 3.10
conda create -n yolov8 python=3.10 -y
conda activate yolov8

# PyTorch (pick CUDA version for your GPU; example: cu121)
pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu121

# Project deps
pip install -r requirements.txt

2) Datasets

• AI-TOD: https://github.com/jwwangchn/AI-TOD
• UAVDT: https://sites.google.com/site/daviddo0323/projects/uavdt
• VisDrone2019-DET: https://github.com/VisDrone/VisDrone-Dataset

Place datasets under datasets/ or edit the YAMLs in ultralytics/cfg/datasets/ (e.g., VisDrone.yaml, ai_tod.yaml, uavdt.yaml).

3) Configs (INR-ready)

• Generic: yolov8_inr_enhanced.yaml
• VisDrone: yolov8_inr_visdrone.yaml
• AI-TOD: yolov8_inr_ai_tod.yaml
• UAVDT: yolov8_inr_uvadt.yaml

All use INREnhancedC2f blocks in backbone/head for tiny-object emphasis.

4) Quick Start (Training)

# Generic INR model
python train.py --model yolov8_inr_enhanced.yaml --data ultralytics/cfg/datasets/VisDrone.yaml --epochs 300

# VisDrone tuned
python train.py --model yolov8_inr_visdrone.yaml --data ultralytics/cfg/datasets/VisDrone.yaml --epochs 300

# AI-TOD tuned
python train.py --model yolov8_inr_ai_tod.yaml --data ultralytics/cfg/datasets/ai_tod.yaml --epochs 300

# UAVDT tuned
python train.py --model yolov8_inr_uvadt.yaml --data ultralytics/cfg/datasets/uavdt.yaml --epochs 300

# Multi-dataset helper script
python train_inr_datasets.py --dataset visdrone   # or ai_tod / uvadt / all

5) Using in Python

import torch
from ultralytics.nn.modules.inr_c2f import INREnhancedC2f

module = INREnhancedC2f(
    c1=256, c2=256, n=3,
    shortcut=False, g=1, e=0.5,
    use_inr=True,       # enable INR enhancement
    use_attention=True, # coordinate-aware attention
    coord_encoding_dim=128
)
x = torch.randn(1, 256, 64, 64)
y = module(x)

6) What is INREnhancedC2f

• Extends C2f with Implicit Neural Representation (INR) to encode fine-grained spatial details for tiny/clustered objects.
• Adds coordinate encoding + attention for better localization under high-altitude UAV views.
• Drop-in replacement for C2f; preserves interfaces so YAML swaps are simple.

7) Key Params to Tune

• use_inr (bool): turn INR enhancement on/off.
• use_attention (bool): keep on for UAV tiny objects.
• coord_encoding_dim (int): 96–192 works well on VisDrone; 128–224 on AI-TOD.
• e (expand ratio) and n (repeats): balance capacity vs. FLOPs for your GPU.

8) Tips for Small-Object UAV Training

• Use higher input resolution if memory allows (e.g., 832–1024 for AI-TOD).
• Keep strong augmentations (mosaic/mixup) modest to avoid over-corrupting tiny targets.
• Validate with class-wise recall; tiny-object recall is the main indicator.

9) References

• INR and coordinate encodings inspired by implicit neural representations for vision.
• Please also cite YOLOv8 and your chosen datasets (AI-TOD, UAVDT, VisDrone) when publishing results.