Feature/#4 implement tl detector.py

[paul-michalik]

OK, the detector is implemented. Tests on synthetic data are green. However, I am unable to run the simulator and the ROS app in sync while using the Windows 10/Docker combo. There seem to be a massive lag in the data transfer - in drive by wire mode the car randomly bumps out of the track after few seconds. Maybe you can run it natively?

[stefancyliax]

I will test it on the local install.

This sounds a lot like the problem Patrick and I had with the waypoint_updater. There the problem was even more pronounced on the local installation. Which is kind of counter intuitive. Problem was that a very minor fault in the waypoint_updater lead to the node publishing very often (depending on system about 280hz) instead of the planned 50hz. So first thing I'll check is rqt_graph..

[paul-michalik]

Hm. The traffic light waypoints are published via /traffic_waypoint based on the frequency of published images and vehicle poses. I haven’t changed the logic of the original implementation at all – but it doesn’t seem to be very robust, IMHO. A /traffic_waypoint is published if a pose and an image were received and if a meaningful traffic light position could be calculated based on this data. From what I can tell, there are no guarantees that these messages match each other in any way…

Sent from my Windows 10 phone

From: stefancyliaxmailto:notifications@github.com Sent: Tuesday, April 3, 2018 9:33 To: stefancyliax/CarND-Capstonemailto:CarND-Capstone@noreply.github.com Cc: Paul Michalikmailto:paul.michalik@outlook.com; Authormailto:author@noreply.github.com Subject: Re: [stefancyliax/CarND-Capstone] Feature/#4 implement tl detector.py (#17)

I will test it on the local install.

This sounds a lot like the problem Patrick and I had with the waypoint_updater. There the problem was even more pronounced on the local installation. Which is kind of counter intuitive. Problem was that a very minor fault in the waypoint_updater lead to the node publishing very often (depending on system about 280hz) instead of the planned 50hz. So first thing I'll check is rqt_graph..

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHubhttps://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fstefancyliax%2FCarND-Capstone%2Fpull%2F17%23issuecomment-378156089&data=02%7C01%7C%7C619b0f57a2eb45424f7908d59935310f%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C636583376082617747&sdata=UF3uuCHz0AeY1WTCFdntJf%2BIlbivEU97zWlWTMXo%2Bgs%3D&reserved=0, or mute the threadhttps://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAKF8bsZdzQlt8k3kVWovQmUFHV4lqRqvks5tkyXEgaJpZM4TEJFN&data=02%7C01%7C%7C619b0f57a2eb45424f7908d59935310f%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C636583376082617747&sdata=u8r1afobsZsCb3QnB%2FYQ2pq4q6pczh7Y2354R4B2%2Fmc%3D&reserved=0.

[stefancyliax]

Understood. So if tl_detector doesn't find any traffic light, nothing is published? But even if, the images should be published at a fixed rate, so this doesn't look like the same problem.

[stefancyliax]

From my first tests, I could affirm that it isn't a problem with publishing "too frequent". It looks like a performance problem with the inference. After I activate the camera images, the CPU load goes to 100% even on the local machine with an i7. Note that I'm using tensorflow on the CPU for now! There is tensorflow-gpu on the machine but its 1.6, which does not work.

So now we have a few options i think:

• install tensorflow-gpu 1.3 (have to downgrade CUDA and cuDNN for this)
• improve performance of inference -> smaller image, smaler net, inference optimizer (TensorflowRT)
• use only every x image

[stefancyliax]

created issues for next steps.

Problems have nothing to do with this branch -> can be merged.