LPIRC CVPR 2019 FAQs

Latest Update: 03/04/2019

 

Online Track

Q : Who can participate?

A : Anyone above 13 years old who is not from US embargoed countries or affiliated with the organizers or the sponsors (Purdue University, Duke University, University of North Carolina Chapel Hill or employees of Facebook or Google Inc).

Q: What phone will be used in the competition?

A : The competition will use Pixel 2. Compared to Pixel 1, Pixel 2's big core has better support for quantized models. You may want to check out the sample latency page that shows timing of the sample models on different devices.

Q : Can Oneline Track use models that are quantized after training?

A : Yes. You are welcome to try but keep in mind that the model's accuracy may be quite different. It's recommended to introduce quantization during training. To save time you can start quantized training from the float model you've obtained. Please see our tutorial page for details

Q : How many times can a team submit?

A : an account is limited to 3 submissions per day. we strongly encourage the team to validate their submissions using the validator before submitting.

Q : Will a leader board be provided?

A : Yes

Q : What does the error messages from the validator mean?

A :

  • "MappedByteBuffer is not a valid flatbuffer model" = Your model is not in TfLite format.
  • "Input error: Failed to get input dimensions. 0-th input should have ... bytes" = The input type should be QUANTIZED_UINT8. Make sure your model is created using the TOCO command listed here.


  • Onsite Track

    Q : What is the time limit in Onsite Track?

    A : 10 minutes

    Q : How many images in Onsite Track?

    A : 20,000

    Q : If a team finished 20,000 images, how can the team reduce energy consumption?

    A : You can reduce the clock frequency to reduce power consumption. You can log out from the server and the power meter stops.

    Q : Do we need to retrieve all images before processing?

    A : You decide the sequence. You can retrieve all images before processing. You can retrieve one image and process one image. You can retrieve multiple images and process them.

    Additional note:

    If you have any questions, do not hesitate to contact Prof. Chen (yiran.chen@duke.edu) or Jingchi Zhang (jingchi.zhang@duke.edu).





    General FAQ

    Q : How is the score calculated in Onsite Track?

    A : mAP and energy consumption

    Q : When can I submit?

    Q :(Extended) To be decided.

    Q : Do we use HTTP protocols? (Onsite Track)

    A : Yes

    Q : Do we need to retrieve all images before processing?

    A : You decide the sequence. You can retrieve all images before processing. You can retrieve one image and process one image. You can retrieve multiple images and process them.

    Q : Is there an advantage of faster processing (more images in 10 minutes)?

    A : A faster solution will likely receive a higher score. Yes, you should make your solution fast.

    Q : What happened with my models that were submitted before May 25?

    A : it is recommended that you resubmit those models.