So I went to update an old project the other day and discovered that MMDetection, the object detection toolbox I have been casually relying on for years, is basically dead. Not sleeping. Not resting. Dead.
And I had no idea.
How I Found Out (The Sad GitHub Way)
I was poking around the repo wondering why there had not been a release in over a year. No bug fixes. No new models. Just silence. So I did what any reasonable person does: I opened the GitHub discussions to see if anyone else had noticed.
Turns out, a lot of people noticed. One user spelled it out pretty plainly in June 2024. OpenMMLab had a steep learning curve, they said, but once you got past it, it was an amazing tool. Then they asked the question nobody wanted to answer: is development stopping?
The response was brutal. Someone explained that MMLab's head professor had passed away, that the team got shuffled to other projects like InternLM, and that even InternLM was barely hanging on.
Half of that held up. The professor part is true and I will get to it. The InternLM part is not: it is Shanghai AI Laboratory's flagship LLM line and it never stopped shipping. InternLM3-8B-Instruct landed in January 2025 and the org's model page has releases from this year. So that was one frustrated user reading the tea leaves, not an obituary. But the frustration about MMDetection specifically was well founded.
I sat there staring at my screen like I had just found out a favorite neighborhood restaurant closed without warning.
What Actually Happened
MMDetection was built by OpenMMLab, the open source ecosystem that grew out of MMLab, the Multimedia Laboratory at the Chinese University of Hong Kong. It was one of those rare open source projects that actually felt polished: modular components, state of the art models, a clean configuration system. You could swap out backbones, necks, heads, and loss functions like Lego bricks. It supported object detection, instance segmentation, panoptic segmentation. The model zoo was ridiculous.
But the whole thing was held together by a key person. Professor Tang Xiaoou. He founded MMLab at CUHK in 2001 and was the engine behind the OpenMMLab ecosystem. When he died on 15 December 2023, the project lost its backbone. Metaphorically and organizationally.
After that, the team was gradually reassigned, and you can watch it happen in the commit log. The last release, v3.3.0, shipped on 5 January 2024. The last commits on main are from 5 February 2024, weights for MM-GroundingDINO. The dev-3.x branch limped on to a single mmcv version bump in July 2024 and then stopped too. The repo is still up (open source can't be taken down), but nobody is home.
What This Means for People Like Me (People Who Cut Corners With Model Zoos)
I have been using MMDetection in projects for years. Training object detectors. Running inference pipelines. The whole workflow was so smooth that I stopped questioning whether the framework would be around next year.
Spoiler: it was not.
Now I am sitting on code that depends on an unmaintained framework. The dependencies (mmcv, mmengine) are also stale. If PyTorch updates something breaking, there is nobody left to patch it. Security issues? Good luck.
I have two options:
- Freeze everything and pray. Pin versions, containerize, never touch it again. This is the path of least resistance and maximum future pain.
- Migrate. Rewrite the detection pipelines using something that is actually maintained. The usual suspects are Detectron2 (Meta), YOLO (Ultralytics), or going direct to PyTorch with torchvision's detection models.
Option 1 is very tempting. Option 2 is what I will probably do after I finish procrastinating.
The Uncomfortable Truth
Open source frameworks with a single institutional backer are fragile. When the key person leaves or passes away, the project often dies quietly. Nobody announces it. There is no press release. The repo just stops moving, and you find out when your CI breaks two years later.
MMLab's other projects are in the same boat. MMSegmentation, MMClassification, MMPose. All built on the same infrastructure, all depending on the same team, all probably headed the same way.
I am not writing this to dunk on anyone. Professor Tang Xiaoou's contributions to computer vision were significant. The MMLab team built tools that made a lot of research and engineering work possible. It is genuinely sad that the projects did not outlast the organization that created them.
But I am also writing this as a note to myself. The next time I build something on top of a great framework, I should ask: what happens if this team disappears tomorrow?
Where That Leaves Us
The repo is still on GitHub. The code still works, for now. The model zoo is still accessible if you know where to look. But nothing new is coming. No bug fixes. No PyTorch compatibility patches. No new architectures.
If you are starting a new project today, MMDetection is not the right choice. Pick something with a pulse. If you have existing MMDetection code, start planning your migration now, while you still remember how everything fits together.
I am going to go update my resume to include "mourns deprecated frameworks" as a skill.