Steep feature dropoff every maxFeatureAge

[violetteavi]

Description:
I was examining the Jupyter notebooks in Kimera-Evaluation when I noticed something strange. The feature count hovers around 300, then has periodic dips to as low as 100 features. After doing a little sleuthing, I discovered that these dips happen every maxFeatureAge number of keyframes, regardless of other factors. I suspect that when you initialize many features at once (like on startup or after a period of motion blur), they all expire at once and must be replaced. I suspect that the only features still being tracked at that point are ones generated from motion-- old features falling out of frame and being replaced by new ones.

Command:

roslaunch kimera_vio_ros <offline launch file>
jupyter notebook

Additional files:
Here are the feature graphs for a maxFeatureAge of 25, varying the intra_keyframe_time. Note that the periodic spikes always coincide with a multiple of 25 keyframes.
intra_keyframe_time = 0.2

intra_keyframe_time = 0.12

intra_keyframe_time = 0.1

intra_keyframe_time = 0.05

Console output:
I also ran the same inputs, but online. Here's a sample of the timing output. Note the huge disparity between max and average time-- I suspect this is because every maxFeatureAge of keyframes, a large number of features die off and must be recomputed.

-----------                        #	Hz	(avg     +- std    )	[min,max]
Data Provider [ms]            	      0	
IMU Preintegration Timing [us]	    661	29.0218	(70.9395 +- 31.0521)	[25,400]
Mesher [ms]                   	     56	2.48127	(0 +- 0)	[0,0]
VioBackEnd [ms]               	     56	2.48136	(48.125 +- 32.0806)	[1,154]
VioFrontEnd [ms]              	    660	28.8361	(27.3697 +- 29.9491)	[5,163]

Please give also the following information:

• SparkVio branch, tag or commit used: master
• SparkVioRos branch, tag or commit used: sequential rosbag
• GTSAM version used:
• OpenGV version used:
• OpenCV version used:
• Operating system and version (e.g. Ubuntu 16.04 or Windows 10): Ubuntu 18.04
• Did you change the source code? (yes / no): yes, extra launchfile and for online changed publishing rate

[violetteavi]

One possible solution would be to cap the number of new features added per keyframe. This would smooth out the dropoff across several keyframes.