Different duration compared to metavision_studio

[fedebecat]

Hi,

I am not sure if I am doing something wrong, I took a look around the issues but I didn't find anything similar.
I have a raw video that I recorded myself using the metavision library. The file is a raw evt3 file.

If I read the file by doing:

wizard = Wizard(encoding="evt3") arr = wizard.read(raw_file) print(f"Recording duration: {(arr[-1]['t']-arr[0]['t'])/int(1e6)} s.")

I get that the duration is 5.14747s, whereas when I open the video with metavision studio I see that the actual duration is 2.2580322s. Any clue about what could be going wrong?
I also recorded the same scene with an RGB camera which should be roughly synchronized with the event camera and I got slightly more than 2 seconds also for RGB.

Thank you!

Federico

[fabrizio-ottati]

Ciao Federico,
yeah, I know about this issue and I have to look into it. It does not happen with the samples on Prophesee website, apart from the EVT3 one. Strange things happen with the timestamps. I will look into it!

[fedebecat]

Ok, thank you very much!

[fedebecat]

If it can be of any use, I compared the timestamps obtained using metavision and the ones obtained with expelliarmus. I noticed that the 12 least significant bits are always the same. I figure that if there is something wrong it should be in the EVT_TIME_HIGH part of the timestamp.

[fabrizio-ottati]

If you would like to try and tackle it yourself, feel free to to do it :) I started another project called misha last year, with the intention of using new pipelines for numpy in C++, but then the last year of PhD happened :) I still want to tackle it though

[lbristiel-psee]

Hi,

I am affraid that there is a problem in the EVT3 decoder code of expelliarmus.
I tested the Wizard on pedestrians.raw (mentioned in the README) and compared to what our EV3 decoder sample outputs (see https://docs.prophesee.ai/stable/samples/standalone/evt3_decoder.html) and events are properly decoded in the begining (first events) but after a while the timestamps starts to diverge and the value given by expelliarmus are too large. So for pedestrians I end up witha length of 65s instead of 30s though the events coordinate are correct.

Here is what I get with expelliarmus

$ python3 test.py 
First event encoded as (t, x, y, p): (5840504, 707, 297, 0)
13th event encoded as (t, x, y, p): (5840587, 689, 302, 0)
130th event encoded as (t, x, y, p): (5845610, 1005, 312, 1)
100000th event encoded as (t, x, y, p): (7143405, 3, 548, 0)
Last event encoded as (t, x, y, p): (65942920, 211, 334, 0)
Number of events: 39297796.
Recording duration: 60.102416 s.

and here is what I get with MV SDK evt3_decoder:

$ sed -n '13p' out.csv 
689,302,0,5840587
$ sed -n '130p' out.csv 
1005,312,1,5841514
$ sed -n '100000p' out.csv 
3,548,0,6492141
$ tail out.csv
211,334,0,35890568

=> same coordinate but not same timestamps.

so someone should double check the source code of EV3 decoded in expelliarmus.
You can use the code of our decoder as an example: https://github.com/prophesee-ai/openeb/blob/main/standalone_samples/metavision_evt3_raw_file_decoder/metavision_evt3_raw_file_decoder.cpp

Hope this helps,
Laurent for Prophesee Support

[fabrizio-ottati]

Nice, thanks @lbristiel-psee! At this point, it would make much more sense to make a lightweight decoder around openeb than maintaining a separate project.

EDIT

This line could be why the durations do not match :) It is not mentioned in the documentation, though.