[QUESTION] Validate the ability of AirDet

[KudoKhang]

@Jaraxxus-Me
First of all, I’d like to thank you for your great work. With this approach, we can address FSOD problems with minimal effort — it’s truly impressive 🔥

Currently, I’m using AirDet in my project (detecting small objects from drone images using three target samples). Based on the results in your paper, I expected AirDet to perform well, but in my case, the object could not be detected.
Could you please share your insights on why it might not work, or point out any mistakes I might have made while using AirDet? Many thanks in advance!

Here’s the input for my case:

Target images: I have three target images with a resolution of 3000×3000, as shown below:

Detection image: The image captured by the drone is low-resolution and slightly blurred, with a dimension of 1024×576:

This project is quite urgent for me, so I’m looking forward to hearing from you soon 🙏🏻

[Jaraxxus-Me]

Hi, Thanks for the interest.
Detecting small objects is a very challenging problem for systems trained on smaller/custom datasets (COCO), such as AirDet (before the foundation model era). Now, given the trend and success of visual foundation models, I would try SAM2 + DINOv3 feature matching in your problem, that will probably be more performant due to their super large pre-training dataset.

[KudoKhang]

Thank you for your quick reply — I really appreciate your explanation.

Since my solution needs to run on an edge device (Jetson Nano), I have to consider both accuracy and computational efficiency when choosing a model. Over the past two days, I’ve done some research, and so far, AirDet still appears to be the most suitable approach for my use case.

However, based on your advice, I’ll also explore the SAM2 + DINOv3 feature matching direction to see how it performs under these constraints.

If you have any other ideas or suggestions that might help with my problem, please feel free to share them — I’d really appreciate your input. I understand you must be very busy, so any comments or guidance from you would be extremely valuable and deeply appreciated 🙏🏻

If you’re interested, I’d also be happy to share my experimental results once I’ve completed the comparison.