Troubleshooting: Behavior recognition

Issues

▪How do I check that the video file contains artifacts?

▪EthoVision XT finds more instances of a behavior than expected

▪EthoVision XT finds no instances of Grooming

▪I get different results when I use Behavior Recognition in two EthoVision XT versions

How do I check that the video file contains artifacts?

Artifacts due to video compression can reduce performance of Behavior Recognition.

1.Create an experiment and set it to tracking From video file and Activity analysis.

2.In the Detection Settings, open the Activity analysis. Reduce the Activity threshold to the minimum.

3.If the video contains artifacts, you should see purple pixels appearing and disappearing with a regular rhythm.

4.Just to be sure, do a test trial and plot the Activity values in the Integrated Visualization. A pattern resulting from video compression artifacts looks like this:

EthoVision XT finds more instances of a behavior than expected

This may occur due to the low contrast of the subject with the background, and too little contrast within the subject area, that is, too little detail in the fur.

For example, too many Grooming instances are scored when the subject is asleep. This occurs when the image of the subject is overexposed. For example, white rats look pure white. In this case the inner part of the subject area is fairly constant in time, compared with the outline, and that increases the probability of detecting grooming. To solve the issue, close the lens’ aperture or reduce the exposure time. See Adjust camera settings in EthoVision XT

EthoVision XT finds no instances of Grooming

The subjects clearly displays bouts of grooming, however EthoVision XT does not detect it at all. This may happen when the regular shape of the subject viewed from above is more “round” than usual, as it occurs in certain strains or circumstances.

In that case, after clicking Grab in the Define Subject Properties dialog often results in a value of Posture being lower than the minimum required (70 for mice). This means that the contracted shape of the subject during grooming is more likely to be considered as “normal”, therefore grooming is not detected.

Find a video frame where the subject walks and looks a bit stretched, preferably when one can see the contour of the hindlimbs. Then, click Grab. Repeat these steps until the value of Posture obtained is higher than the minimum required.

I get different results when I use Behavior Recognition in two EthoVision XT versions

There may be differences in the scores of behaviors between EthoVision XT 19 and a previous version, even when using the same video file and the same Behavior Recognition settings. This may have to do with the different video decoding software that is used in the two versions. Small changes in the way video is displayed (e.g. the same pixel displayed with slightly different intensity value) may result in changes in the probability of some behaviors, and therefore influence the automatic scores.

We tested a number of video files and found that the codec used to create the video file may affect the difference in the results between software versions. When the video was created with H.264 MPEG-4 AVC codec or the mp4v MPEG-4 codec, the average difference in the behavior scores (either frequency of occurrence or cumulative duration) expressed as a percentage of the scores obtained with the same video in EthoVision XT 15, is between 0 and 2% (averaged across all behavior categories).

The H.264 MPEG-4 AVC codec is used by EthoVision XT 16-18 and MediaRecorder 5-6. The mp4v MPEG-4 codec is used by EthoVision XT 15 and MediaRecorder 4.

For codecs like DivX, Xvid, MPEG 1 and MPEG 2, the average difference between EthoVision XT 15 and 16 is larger - up to 22%.

As general recommendations:

▪Always use EthoVision XT or MediaRecorder to record video that you will analyze with Behavior Recognition.

▪When you upgrade an experiment made with an older version of EthoVision XT to EthoVision XT 19, and you want to re-do Behavior Recognition, make sure you use exactly the same Arena Settings and Detection Settings that you used with that particular video file.

▪In the Analysis profile, adjust the criteria based on the behavior probability to only consider the instances of behaviors that have a high probability, for example higher than 95%. That usually gives more robust results.