discussion / Camera Traps  / 21 November 2023

Video camera trap analysis help


I'm a complete newbie so any help would be appreciated. I never trained any ML models (have junior/mid experience with Python and R) nor annotated the data but would like to learn.


We have set couple of camera traps that record videos on vulture feeding sites and I am tasked to analyze the video data for their presence.

I thought that using Megadetector could work but it seems to me it only takes in images so I don't know where to start.


What would you use, is there a pipeline (or articles/repositories/video) about how best to approach the task?


Thanks in advance.

Hey there :)

You could use Timelapse (for data annotation and automatic metadata extraction, e.g. date/time, temperature, etc) in combination with Megadetector (it works with videos too). You basically run the Megadetector through your batch of videos and then use the output json file in Timelapse (to filter out empty videos or automate the annotation process). 

Here's the documentation for Timelapse:


And here you can find detailed info on how to integrate the Megadetector input into Timelapse.


I hope this is useful.



You are correct that at its core, MegaDetector only knows about still images.  However, we have found that it's quite reasonable to run MegaDetector on something like every 10th frame of a video (that's still typically three frames per second) to determine whether animals are present.  I would say video represents around 10% of total MegaDetector use right now.  If you have very small animals that flash in and out of frame quickly, this probably isn't a good approach (this comes up a lot with video surveys for birds), but for most animals, this has worked great.

As Lucy highlights, most MegaDetector users load their MegaDetector results into Timelapse to review their images; Timelapse lets you say, for example, "show me only the images that MegaDetector thinks are not empty".  It can do the same for videos.  But it's not necessary to use Timelapse; some users, for example, review their videos in dedicated behavioral coding software and just want to get blanks out of the way first.  MegaDetector can help with that.

All that said, it's a little bit more involved to process videos than to process stills, so we recommend you reach out to us at [email protected], and we can run a few test videos, make sure things are working OK, then walk you through the process.  For the adventerous, the relevant script is here, and the notebook we use to do this for batch jobs is here, but for new users, we typically recommend starting by email.

And just for fun, here is a video of MegaDetector running on a video of a raccoon.  Because raccoons.