Meteorix DETR-ResNet-50 (meteor detection)

DETR-ResNet-50 fine-tuned from facebook/detr-resnet-50 to detect meteors (single class meteor) in ground-based night-sky images, as part of the Meteorix study comparing YOLO26n, RF-DETR Small and DETR-ResNet-50.

Usage

import torch
from PIL import Image
from transformers import DetrForObjectDetection, DetrImageProcessor

model = DetrForObjectDetection.from_pretrained("Didrien/meteorix-detr-resnet50").eval()
processor = DetrImageProcessor.from_pretrained("Didrien/meteorix-detr-resnet50")

image = Image.open("sky.jpg").convert("RGB")
width, height = image.size
# The model was trained on square 800x800 images only: resize to that format before inference.
inputs = processor(images=image.resize((800, 800)), return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
detections = processor.post_process_object_detection(
    outputs, threshold=0.5, target_sizes=torch.tensor([[height, width]])
)[0]

Feed square images: on non-square images given at their native aspect ratio, the predicted boxes are shifted.

Training

• Full fine-tuning of every layer, 300 epochs, AdamW (lr 1e-4, backbone 1e-5, weight decay 1e-4), learning rate divided by 10 after epoch 200, effective batch size 64, gradient clipping 0.1, horizontal flip and multi-scale resizing (shortest side 480-800 px), auxiliary decoder losses.
• Data: Mendeley meteor image dataset (Gradvohl & Shirasuna, DOI 10.17632/4k98n84d9g.1, CC BY 4.0), 3,849 training / 550 validation / 1,100 test images, 800x800.

Results (image level, confidence 0.5, matching IoU 0.5)

Mean latency on an RTX 4090: 21.6 ms per image.