Why tiny humans are hard to find in thermal images

Vidhan Jain — Mon, 20 Jul 2026 00:00:00 +0000

For two years, my job was to find people in images where a person is sometimes smaller than the cursor you're reading this with. Here's why that's harder than it sounds, and what kind of ideas help.

The setup

A drone flies over a field, a road, or a forest at night with a thermal (long-wave infrared) camera. Warm things glow; cool things don't. From high up, a standing person might cover a patch of just a few pixels. The task: draw a box around every person.

Benchmarks like TinyPerson exist precisely because this is where normal detectors fall apart.

Why it's hard

There's almost nothing to see. A detector recognises a person by shape and texture: head, shoulders, legs. At a few pixels, none of that exists. You get a warm blob, and plenty of other things are warm blobs too: rocks that soaked up sunlight, car engines, animals.

The network throws the evidence away. Most detectors shrink the image step by step (downsampling) to build up high-level features. Each step halves the resolution. A person who was 8 pixels wide at the input is a fraction of a pixel a few layers later, and their signal gets averaged into the background.

Contrast isn't guaranteed. Thermal contrast depends on the temperature difference between the person and their surroundings. On a warm evening, the ground can be almost as warm as a body, and the person nearly disappears.

It has to be light. If the detector is going to run on the drone itself, it can't be a 200 GFLOP model. Every trick you add has to pay for itself.

What helps

Across the four architectures in my thesis, a few ideas kept showing up:

Protect the small signal early. Refine and re-weight features in the early, high-resolution layers, before downsampling blurs the evidence away.
Look at the neighbourhood. A warm blob on a road is more likely a person than a warm blob in a parking lot full of cars. Context from the surrounding area, compared against the centre, helps separate targets from clutter.
Use frequency information. Tiny targets live in the high-frequency part of the image. Treating high and low frequencies differently helps keep them from being smoothed out.
Merge scales carefully. When features from different resolutions are combined, small misalignments hurt tiny objects most. Fusing them with care matters more than stacking more layers.

The details are in the papers on the research page, including TSRNet and SAPNet.

The part I didn't expect

The biggest lesson wasn't architectural. It was that careful evaluation matters as much as the model. On tiny-object benchmarks, small changes in how boxes are matched or which scales are counted can move numbers more than a new module does. I'll write more about that in a future post.

Redesign: A new home for my writing

Vidhan Jain — Wed, 30 Sep 2026 00:00:00 +0000

For a while, vidhanjain.com was a single page: a big gradient headline, a grid of glowing cards, and a "Get in touch" button. It looked fine. It also said almost nothing about me that my resume didn't already say better.

So I rebuilt it as a personal website instead of a portfolio.

What changed

A portfolio is a sales page. It's written for someone deciding whether to hire you, and it's done the moment it's published. A personal site is closer to a notebook that happens to be public. It grows, it has opinions, and it's allowed to be unfinished.

The new site is split into a few pages, each with one job:

Blog for notes and longer posts
Research for papers, each with a summary in plain words
Projects for things I've built
Photos for the non-work side of life
About me for everything else, plus a now page

How it's built

It's a static site built with Eleventy and hosted on Vercel. Blog posts are Markdown files. Research papers and projects each live in a small JSON file, and the photos page pulls in whatever I upload to Pinterest, so adding a paper means adding a few lines instead of copying a block of HTML around.

src/
├── _data/          # publications, projects, photos, site info
├── _includes/      # layouts and shared pieces (sidebar, footer)
├── blog/           # one Markdown file per post
├── assets/         # CSS, JS, images, photos
└── *.njk           # one file per page

No framework, no client-side routing, one stylesheet. Pages load fast and should keep working for a long time.

What's next

Mostly writing. I have a backlog of things I learned during my M.Tech that never made it out of my lab notebook: why tiny objects are so hard for detectors, what actually moved the needle in my ablations, and how I set up experiments in MMDetection. Those are coming.

If you want to follow along, there's an RSS feed.