ATOM and DiMP Learn to Track Online

ATOM and DiMP keep a small target model that updates every few frames, so the tracker adapts to how this specific object looks right now.

ATOM predicts box overlap while DiMP trains a tiny online classifier, lifting the target score from 0.52 to 0.68 in ten update steps

Why Does This Exist?

Fixed siamese templates go stale the moment lighting or pose shifts. The tracker needs to learn what this target looks like now, not what it looked like in frame one. Online learning fills that gap.

ATOM adds an overlap predictor that refines boxes by predicted IoU, and DiMP replaces hand-tuned updates with a learned optimizer for a tiny classifier. Both sit on single-object tracking, reuse the matching idea from siamese networks, and explain why SiamRPN boxes still need refinement under change.

Think of It Like This

A bouncer updating a photo at the door

A club bouncer holds one entry photo, then sketches quick notes as each guest returns in a new jacket or hat. The sketch stays of the same person, just fresher.

It stops working when the bouncer starts sketching the crowd behind the guest. Online updates also absorb background when the target is occluded.

How It Actually Works

Let $f$ be backbone features and $w$ a tiny $1 \times 1$ classifier, where $w$ holds target versus background weights and $f$ holds the current frame embedding. The response is $s = w * f$ , a convolution that should peak on the target. Each update solves a ridge loss $L = \|s - y\|^2 + \lambda\|w\|^2$ , where $y$ is a Gaussian label map and $\lambda$ is weight decay.

Worked example

Start with overlap 0.52 and classifier margin 0.4. Run ten steepest-descent steps on five hard negatives. The peak climbs 0.52, 0.58, 0.62, 0.65, then 0.68 as $w$ sharpens. ATOM reads the 0.68 box as highest predicted IoU and keeps it. The arithmetic is small but the direction is what matters: each step moves $w$ toward the target patch and away from the nearest distractor.

Watch Out For

Updating through full occlusion

Scores dip below 0.3 when the target hides, yet updates keep running and $w$ learns the occluder. The symptom is a confident box on the wrong object after reappearance. Freeze updates when the peak falls under threshold and resume only on a clean high peak.

The Quick Version

ATOM predicts IoU to refine boxes instead of trusting raw similarity peaks.
DiMP learns the online optimizer so a tiny classifier adapts in a few steps.
The fixture lifts overlap from 0.52 to 0.68 in ten updates.
Online models adapt fast but absorb background during occlusion.
Freeze updates on low peaks to keep the model clean.