Dual-Imbalance Continual Learning for Real-World Food Recognition
arXiv:2603.29133v1 Announce Type: new Abstract: Visual food recognition in real-world dietary logging scenarios naturally exhibits severe data imbalance, where a small number of food categories appear frequently while many others occur rarely, resulting in long-tailed class distributions. In practice, food recognition systems often operate in a continual learning setting, where new categories are introduced sequentially over time. However, existing studies typically assume that each incremental step introduces a similar number of new food classes, which rarely happens in real world where the number of newly observed categories can vary significantly across steps, leading to highly uneven learning dynamics. As a result, continual food recognition exhibits a dual imbalance: imbalanced sample
View PDF HTML (experimental)
Abstract:Visual food recognition in real-world dietary logging scenarios naturally exhibits severe data imbalance, where a small number of food categories appear frequently while many others occur rarely, resulting in long-tailed class distributions. In practice, food recognition systems often operate in a continual learning setting, where new categories are introduced sequentially over time. However, existing studies typically assume that each incremental step introduces a similar number of new food classes, which rarely happens in real world where the number of newly observed categories can vary significantly across steps, leading to highly uneven learning dynamics. As a result, continual food recognition exhibits a dual imbalance: imbalanced samples within each food class and imbalanced numbers of new food classes to learn at each incremental learning step. In this work, we introduce DIME, a Dual-Imbalance-aware Adapter Merging framework for continual food recognition. DIME learns lightweight adapters for each task using parameter-efficient fine-tuning and progressively integrates them through a class-count guided spectral merging strategy. A rank-wise threshold modulation mechanism further stabilizes the merging process by preserving dominant knowledge while allowing adaptive updates. The resulting model maintains a single merged adapter for inference, enabling efficient deployment without accumulating task-specific modules. Experiments on realistic long-tailed food benchmarks under our step-imbalanced setup show that the proposed method consistently improves by more than 3% over the strongest existing continual learning baselines. Code is available at this https URL.
Comments: Accepted to 3rd MetaFood at CVPR 2026. Code is available at this https URL
Subjects:
Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2603.29133 [cs.CV]
(or arXiv:2603.29133v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2603.29133
arXiv-issued DOI via DataCite (pending registration)
Submission history
From: Jiangpeng He [view email] [v1] Tue, 31 Mar 2026 01:32:51 UTC (4,788 KB)
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
modelbenchmarkannounce
Thursday: April 2 - AI, ML and Computer Vision Meetup
<p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F13z7vsfxskklxi6zis2w.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F13z7vsfxskklxi6zis2w.png" alt=" " width="800" height="420"></a></p> <p>Join us on April 2 at 9 AM Pacific for the monthly AI, ML and Computer Vision Meetup!</p> <p><a href="https://voxel51.com/events/ai-ml-and-computer-vision-meetup-april-2-2026" rel="noopener noreferrer"><strong>Register for the Zoom</strong></a></p> <p>Talks will include:</p> <ul> <li> <strong>Visual AI
BuzzFeed Is Dying Because It Bet Everything on AI — And Its CEO Still Won't Admit It
<p>BuzzFeed is circling the drain. The company just reported a net loss of $57.3 million for 2025, its stock is trading at $0.70 — that's not a typo, seventy cents — and in its latest earnings report, the company included the phrase every investor dreads: "substantial doubt about the Company's ability to continue as a going concern."</p> <p>Translation: we might go bankrupt.</p> <p>And the man responsible for this spectacular implosion, CEO Jonah Peretti, has learned absolutely nothing. His plan? Build more AI apps.</p> <p>You cannot make this up.</p> <h2> The Anatomy of a Self-Inflicted Wound </h2> <p>Rewind to January 2023. ChatGPT had just exploded into public consciousness. Every CEO with a quarterly earnings call was scrambling to say "AI" as many times as humanly possible. Peretti wa
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Releases
Why Your AI Agent Health Check Is Lying to You
<p>Your monitoring dashboard shows green across the board. Process running. Port responding. CPU normal. Memory stable.</p> <p>But your AI agent hasn't done anything useful in four hours.</p> <h2> The problem with traditional health checks </h2> <p>Traditional health checks answer one question: "Is the process alive?" For web servers, that's usually enough. If Nginx is running and responding on port 80, it's probably serving pages.</p> <p>AI agents are different. An agent can be alive without being productive. The process is running, but the main work loop is stuck on a hung HTTP call, waiting on a deadlocked mutex, or spinning in a retry loop that will never succeed.</p> <h2> Three ways health checks lie </h2> <h3> 1. PID exists ≠ working </h3> <p><code>systemctl status my-agent</code> sa
Building an Engineering & Security News Aggregator (10 Sources, No APIs)
<p>We built a curated engineering and security news aggregator that pulls from 10 high-signal sources, deduplicates content, and updates every 6 hours.</p> <p>No paid APIs. No scraping. No login. Just clean, structured news for developers.</p> <p>This post breaks down exactly how it works.</p> <h2> What This Is </h2> <p>A lightweight news wire combining:</p> <ul> <li>Hacker News </li> <li>Lobsters </li> <li>InfoQ </li> <li>Cloudflare Blog </li> <li>Krebs on Security </li> <li>The Hacker News (Security) </li> <li>NIST NVD (vulnerabilities) </li> <li>GitHub Blog </li> <li>OpenAI Blog </li> <li>Anthropic Research </li> </ul> <p>The goal: <strong>high-quality signal, zero noise, zero cost</strong>.</p> <h2> Why Build This? </h2> <p>Most engineering/news aggregators fail in one of these ways:</
How I Pick Crypto Trading Pairs for My Bot — A Data-Driven Framework
<p>Choosing the right trading pairs is one of the most underrated aspects of building a profitable crypto bot. After months of testing, here's the framework I use to select and rotate pairs.</p> <h2> Why Pair Selection Matters </h2> <p>Most tutorials focus on entry signals and indicators. But trading the wrong pairs can kill even a great strategy. A perfect RSI reversal signal on a low-liquidity altcoin will get eaten by spread and slippage.</p> <p>My bot trades 15 pairs on Bybit futures. Here's how I picked them.</p> <h2> The 5 Criteria I Use </h2> <h3> 1. Minimum Daily Volume: $50M+ </h3> <p>Anything below $50M in 24h volume means:</p> <ul> <li>Wide spreads that eat your profits</li> <li>Slippage on entries and exits</li> <li>Gaps that trigger false signals</li> </ul> <p>I check volume o
Telepage – I built a self-hosted PHP app that turns any Telegram channel into a website
<p>If you run a Telegram channel, you already know the problem: your content is invisible to Google, there's no search, old posts are buried, and readers need the app just to see your work.</p> <p>I built <strong>Telepage</strong> to fix that.</p> <h2> What it does </h2> <p>Telepage connects to your Telegram channel via a bot webhook and turns every post into a searchable web card — automatically, in real time.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Ft0zcxkaob0c3c5sglju1.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uplo
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!