You Only Stack Once (YOSO): A Motion-Filtered, Deep-Learning Framework for Detecting Faint Moving Sources

The paper introduces You Only Stack Once (YOSO), a novel deep-learning pipeline that utilizes a Gaussian Motion Filter to efficiently detect faint, slow-moving Solar System objects with an extremely low false positive rate, offering a scalable alternative to traditional shift-and-stack methods for large-scale astronomical surveys.

The Gist

Imagine you are trying to find a tiny, slow-moving firefly in a vast, dark field. The problem is that the firefly is so dim that in any single snapshot you take with your camera, it's invisible. It's just a speck of dust. However, if you take 100 photos in a row, that firefly moves a tiny bit in each one, leaving a faint, broken trail.

For decades, astronomers have tried to find these "space fireflies" (like distant icy rocks called Trans-Neptunian Objects) using a method called "Shift-and-Stack."

The Old Way: The "Guess-and-Check" Game

The traditional method is like trying to line up 100 photos of that firefly by guessing how fast it's moving.

1. You guess: "Maybe it's moving 1 inch per second." You shift the photos to match that speed and stack them. If the firefly appears, great!
2. If not, you guess: "Maybe it's moving 1.1 inches per second." You shift and stack again.
3. You keep doing this for every possible speed and direction.

The Problem: This is like trying to find a needle in a haystack by building a million different haystacks. Because you are trying so many different speeds, you often accidentally line up random dust or noise in a way that looks like a firefly. This creates "false alarms" (false positives). To fix this, astronomers have to manually check thousands of these fake fireflies, which takes forever.

The New Way: "You Only Stack Once" (YOSO)

The paper introduces a new system called YOSO (You Only Stack Once). Instead of guessing the speed and trying a million different ways to stack the photos, YOSO uses a clever trick and a smart computer brain (Artificial Intelligence).

Step 1: The "Motion Filter" (The Magic Lens)

Imagine you have a special filter that only highlights things that are moving in a specific way, while ignoring everything else.

• How it works: The paper uses a "Gaussian Motion Filter." Think of this as a mathematical lens that looks at every single pixel in your photos over time.
• The Analogy: If a star is just sitting still, it looks like a steady dot. If a firefly flies past, it creates a specific "pulse" of light as it enters and leaves a pixel. The filter amplifies that specific pulse and smoothes out the random static noise.
• The Result: Instead of trying to line up the photos perfectly, the filter turns the broken trail of the moving object into a single, bright, continuous line in one combined image. You only have to combine the photos once.

Step 2: The "Smart Detective" (YOLOv8)

Once the photos are combined into this single image with bright trails, a computer program (based on a system called YOLOv8, which is famous for spotting objects in real-time video) scans the image.

• The Analogy: Think of this AI as a highly trained dog that has been shown thousands of pictures of "space firefly trails" and "fake noise." It instantly sniffs out the real trails and ignores the dust.
• The Benefit: Because the AI is looking for a specific shape (a trail) rather than just a bright dot, it makes very few mistakes. The paper claims the "false alarm" rate is incredibly low.

Step 3: The "Fine-Tuning" (Double-Check)

When the AI spots a trail, the system does one final, quick check. It takes that specific trail and runs a tiny, focused version of the old "shift-and-stack" method just for that one object. This confirms the exact speed and direction, turning the trail back into a sharp, round dot so astronomers can measure how bright it is.

What Did They Find?

The team tested this new system on data from the Dark Energy Camera (DECam), looking at a patch of sky where they already knew some objects were hiding.

• The Catch: The new system wasn't quite as good at finding the very faintest objects as the old "guess-and-check" method (it missed the dimmest ones).
• The Win: However, it was much faster and had far fewer false alarms.
• The Discovery: Even though it was "shallower" (didn't see the dimmest things), it found 11 new objects that the old method missed! It also found 216 fast-moving objects (like asteroids) that the old method wasn't even looking for.

Why Does This Matter?

The paper argues that this method is a game-changer for the future of astronomy, specifically for the Large Synoptic Survey Telescope (LSST), which will take millions of photos of the sky every night.

• Efficiency: Instead of spending years trying to guess the speed of every object, LSST can use YOSO to process data instantly.
• Versatility: The paper suggests this same "motion filter" idea could be used for other things, like finding planets around other stars (by looking for their tiny wobbles) or spotting fast-moving space rocks that could hit Earth.

In short: YOSO stops trying to guess the speed of the universe and instead uses a smart filter and a computer brain to spot the trails left behind, making the search for hidden space rocks faster, cleaner, and surprisingly effective.

Technical Summary

Technical Summary: You Only Stack Once (YOSO)

Problem Statement
Detecting faint, slow-moving Solar System objects, particularly Trans-Neptunian Objects (TNOs), remains a significant challenge in wide-field astronomical surveys due to their low surface brightness (often $m > 22$). Traditional detection relies on the "shift-and-stack" method, which aligns images based on trial-and-error velocity vectors to compensate for parallactic motion. While effective, this approach suffers from two primary limitations: (1) the computational cost scales with the number of required velocity and direction trials, and (2) the accumulation of false positives increases as the number of trials grows, often misidentifying static background features as moving sources. Furthermore, existing deep learning applications in astronomy have largely focused on static classification or specific anomaly detection, leaving a gap for motion-based signal enhancement that does not rely on exhaustive velocity grids.

Methodology
The authors present You Only Stack Once (YOSO), an automated pipeline that integrates a novel Gaussian Motion Filter (GMoF) with a Convolutional Neural Network (CNN) to detect moving sources without the need for discrete velocity trials.

1. Pre-processing and Image Subtraction: The pipeline utilizes DECam InstCal images. It performs local background subtraction using photutils and applies the ISIS package (Alard–Lupton optimal image subtraction) to remove stationary sources, leaving difference images containing only moving or varying objects.
2. Gaussian Motion Filter (GMoF): Instead of shifting images to a specific rate, YOSO operates on the pixel level. It recognizes that a moving point source traversing a pixel creates a light curve that approximates a Gaussian function. The GMoF applies a one-dimensional Gaussian filter ($G_{mov}$) to the time-series of each pixel. The filter width ($\sigma_{filt}$) is derived from the object's apparent velocity, the Point Spread Function (FWHM), and the exposure cadence. The detection statistic for each pixel is defined as the maximum value of the convolved light curve ($max(I(t) * G_{mov})$). This process generates a single combined image where moving sources appear as enhanced linear trails, while random noise and static residuals are suppressed.
3. Deep Learning Detection: The resulting motion-filtered stack is processed by a YOLOv8-L (You Only Look Once) object detection model. The model was trained on a synthetic dataset of over 68,000 objects implanted into DECam fields, covering magnitudes $m_{VR} \in [19, 27]$ and apparent velocities of 3–5 arcsec/hr. The training data included augmentation for brightness, contrast, and PSF variations to ensure robustness.
4. Refinement and Vetting: While the CNN identifies candidate trails, it does not uniquely determine the direction or precise rate of motion. To resolve this, the pipeline performs a bi-directional shift-and-stack refinement on the candidate regions. It uses the Source Extraction and Photometry (SEP) library to measure source ellipticity. The correct motion vector is identified by minimizing the ellipticity of the stacked source (aiming for a point source) while maximizing flux. Candidates are retained only if they meet strict criteria: detection above $1.5\sigma$, circular morphology ($FWHM \ge 2.5$), and very low ellipticity ($e < 0.001$).

Key Results
The method was applied to a subset of the DECam Ecliptic Exploration Project (DEEP) survey, specifically four field nights (FNs) in the B1 quadrant.

• Recovery Efficiency: YOSO successfully re-detected 45 out of 73 previously identified TNOs from the DEEP collaboration's KBMOD analysis.
• New Discoveries: The pipeline identified 11 new TNOs and 216 previously unreported faster-moving objects (apparent rates 10–60 arcsec/hr), including Near-Earth Objects (NEOs) and Main-Belt asteroids.
• Limiting Magnitude: The method reaches a limiting magnitude approximately 0.88 magnitudes brighter than the KBMOD shift-and-stack method. However, the authors note that YOSO's false positive rate is extremely low because it requires a source to exhibit a trail and subsequently resolve into a point source at the correct rate.
• Performance Metrics: On the validation set, the model achieved a precision of 0.80 and a recall of 0.71. The final vetting process yields a purity rate close to 99%.
• Computational Efficiency: Unlike shift-and-stack methods that require a dense grid of trials, YOSO generates a single combined image and performs inference in under 11 milliseconds per image, making it highly scalable.

Significance and Claims
The paper claims that YOSO provides a versatile, scalable alternative to traditional moving object detection, particularly for data-intensive astronomy. Its primary significance lies in:

• False Positive Suppression: By detecting only sources that form trails and resolve into point sources, the method drastically reduces the false positive rate inherent in high-volume trial-and-error stacking.
• Scalability: The "stack once" approach reduces computational overhead, making it suitable for large-scale surveys like the LSST (Large Synoptic Survey Telescope), particularly in Deep Drilling Fields where intra-night sampling is dense.
• Broad Applicability: The authors demonstrate that the framework is not limited to slow-moving TNOs but effectively detects fast-moving NEOs and Main-Belt asteroids using a single velocity hypothesis.
• Adaptability: The paper suggests the GMoF technique can be adapted for other domains requiring motion-based signal enhancement, including:
  • Exoplanet Imaging: Adapting the method for Angular Differential Imaging (ADI) to distinguish planetary signals from quasi-static speckles.
  • Space Missions: Potential onboard processing for missions like NEO Surveyor to reduce data transmission volume by transmitting stacked frames rather than raw exposures.
  • Time-Domain Astronomy: Identifying subtle brightness variations in variable stars or supernova progenitors.

The authors conclude that while YOSO currently does not reach the same depth as KBMOD for the faintest objects, its high purity, computational efficiency, and ability to recover objects in complex backgrounds make it a powerful complementary tool for Solar System science and future wide-field surveys.