One of our cameras started doing something odd. Someone would walk through the frame, and after they moved on, a dark, person-shaped smear stayed behind where they had been for a second or so before it faded out. It looked like the picture couldn’t keep up.
We went straight for the usual suspects: the stream, the network, the phone. We even set up a test that throttled a client on purpose and got go2rtc to drop packets, and it felt like proof. It wasn’t. The real cause was a setting on the camera itself, and the clue that would have saved us an afternoon was in the first five minutes.
What a ghost trail looks like
It’s worth being precise, because a few different problems all get called “smearing”.
A network or decode problem usually looks broken. You get grey or green blocks, parts of the picture that freeze while other parts move, or a stall followed by a jump forward. It tends to hit a whole frame or a whole slice of one, and it often shows up on one viewer and not another.
A noise reduction ghost looks smooth. The moving thing leaves a soft, darker, half-transparent copy of itself where it just was, and the copy fades over a second or so. The rest of the picture is clean. It’s worst in low light, and it shows up when the camera has switched to black-and-white night mode, even in the middle of the day if the scene is dim enough.
The checks that actually pointed at the camera
Only one camera did it. Every other camera on the same server, same network and same app looked fine. That alone should have stopped the network theory. If the path between the server and the phone were the problem, the other cameras would have had it too.
It wasn’t the busiest stream. The affected camera turned out to have the lowest bitrate of the lot. Bandwidth problems hit the heavy streams first, not the lightest one.
The settings matched a healthy camera. Its encoder and image settings were the same as a camera that had no ghosting at all. The only real difference was the scene: a dim view where the camera sat in night mode a lot, versus a well-lit one.
It was in the recording. This is the one that settles it. We pulled the recorded segment off the server and stepped through it frame by frame. The ghost was there in the file. A recording is written straight from the camera’s stream on the server, with no phone, no remote connection and no player involved, so anything visible in the saved file happened before the video ever left the camera.
If you take one thing from this post, make it that last check. Look at the recording, or the camera’s own web page, before you change anything on the network.
Why noise reduction does this
Small camera sensors produce a lot of grain in low light. To hide it, most cameras run temporal noise reduction, often labelled 3D noise reduction or 3D DNR in the settings. It works by blending each frame with the ones before it. Grain is random from frame to frame, so it averages away. A wall that isn’t moving stays put, so it comes through clean.
The trade-off is motion. When something moves, the blend still carries a bit of where it was a moment ago. In good light the camera can tell motion from noise easily and keeps the blending light. In a dim room, especially on the black-and-white night image, almost everything looks like noise to it, so it blends harder, and moving objects leave a trail.
The fix, and what it costs
We turned 3D noise reduction off on that one camera. It’s usually in the camera’s image or ISP settings, next to things like day/night mode and exposure. Some cameras give you a strength level instead of just on or off, and if yours does, try a lower level first.
The ghosting stopped. People moving through the frame stay sharp now.
What you give up is the grain filter. Night footage on that camera is noticeably grainier than it was. Grain also isn’t free to encode: the camera has to spend bits describing noise that used to be smoothed away, so at a fixed bitrate the rest of the picture gets a bit less detail. We haven’t seen it hurt detection so far, but we’re watching the night footage before we call it settled, and we left noise reduction on for every camera that wasn’t ghosting.
If turning it off makes the grain worse than the trails were, the other fix is light. More light on the scene means the camera stays in color mode longer and needs less noise reduction to begin with. A lamp on a timer or an extra IR illuminator can do more for a dim room than any setting.
What we’d do next time
Before blaming the network for any odd-looking video, we’d now ask three questions in this order. Is it only one camera? Is it in the recording? And what is different about that camera’s scene, not its settings? On this one, those three answers were there from the start. We just went looking somewhere more interesting first.