PiattiDetector — Aerial Object Detection (SkyFusion)

Lightweight anchor-free object detector for aerial imagery. Detects aircraft, ships and vehicles using a CenterNet-inspired approach with heatmap + size + offset regression heads.

Developed as part of a Machine Learning course (Sapienza, January 2026).
Code: github.com/JavideuS/aerial-vision · regression/

Architecture

Custom ResNet backbone (3 residual blocks, 8× downsampling) + three detection heads:

Head	Output	Purpose
Heatmap	`[B, C, H/8, W/8]`	Object center confidence per class
Size	`[B, 2, H/8, W/8]`	Bounding box width and height
Offset	`[B, 2, H/8, W/8]`	Sub-pixel center refinement

~2.1M parameters. Input: 512×512 px.

Dataset

SkyFusion / Tiny Object Detection — 3000 aerial images, 640×640 px, COCO format.
Classes: Aircraft, Ship, Vehicle.

Usage

from huggingface_hub import hf_hub_download
import torch
from modelv2 import PiattiObjectDetector  # clone the repo first

ckpt_path = hf_hub_download(
    repo_id="JavideuS/skyfusion-flood-reg",
    filename="PiattiDetector-2111k.pth"
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = PiattiObjectDetector(num_classes=3).to(device)
ckpt = torch.load(ckpt_path, map_location=device)
model.load_state_dict(ckpt['model_state_dict'])
model.eval()

For full inference with bounding box decoding and visualization see main.py.

Limitations

Trained on a small dataset (3000 images); may not generalise to aerial imagery with significantly different sensors or altitudes
No NMS post-processing in the provided inference script

Downloads last month: -; Downloads are not tracked for this model. How to track