Submission instructions and reporting guidelines

The proponents of all registered proposals will receive by email a TEAMID registration ID with (two digits) along with login and password to upload all the required information.

File submission: data organization

There should be four seperate ZIP files with the following information:

  • Decoder+docker file: source decoder and docker configuration file. If justified, the Docker container could be considered for submission.
  • Model(s) to be used during decoding. The following naming convention must be used: <TEAMID>_<BR>.model. If only one model is used <BR> should be set to ‘all’.
  • Decoded images according to the naming convention defined bellow.
  • Bitstreams to be used for decoding using the same naming convention as the decoded images but with .bits extension.

The naming convention for the zip files is TEAMID_DATETIME_<[Decoder_docker | Models | Bitstream | Decoded]>.zip where DATETIME is MMDDHHMM, e.g. the filename for team 01 to submit the decoder models in 12th June at 10:54 (CEST) is 01_12061054_Models.zip.

Test images

The test images will be available here. See when in Timeline.

Docker file

To ensure the proponents computing environment (to run the submitted decoder) can be reproduced, a Dockerfile is required as part of the full submission.

To configure a suitable environment, it is proposed to start from one of the following Docker images:

  • Tensorflow
    • tensorflow/tensorflow:1.15.0-gpu-py3 (GPU)
    • tensorflow/tensorflow:1.15.0-py3 (CPU)
    • tensorflow/tensorflow:2.0.0-gpu-py3 (GPU)
    • tensorflow/tensorflow:2.0.0-py3 (CPU)
  • Pytorch
    • pytorch/pytorch:1.3-cuda10.1-cudnn7-runtime (GPU)

The decoder must be able to run in either one of the above listed Docker images or an image specified on your own. In either case, a Dockerfile that allows for replication of your environment must be part of the full submission.

For instructions on how to work with Docker we recommend the official documentation.

Memory requirements

Report the amount of memory and GPU model required for decoding (inference) here.

Bitrate reporting and cross-check verification

The following information is needed for all decoded test images:

  • Target and real bitrates.
  • Mean Square Error (MSE) using the corresponding original image as reference.

The proponents should use the following CSV template to be included in the ZIP file with the decoded images submission.

Decoded images naming convention

The decoded files in PNG must be performed according to the following naming convention: <TEAMID>_<IMGID>_TE_<RES>_8bit_sRGB_<BR>.png where:

  • TEAMID is the registration team ID attributed with 2 digits
  • IMGID is an identification of the image with 5 digits
  • TE is a fixed value which represents it is a test image
  • RES is the spatial resolution (width x height)
  • Bit depth (which must be 8 bits always)
  • Color space (which must be sRGB)
  • BR target bitrate for decoded images: YXX (e.g. 1.25 bpp would be ‘125’ and 0.05 would be 005)

File submission: connection details

The submission must be performed by using sftp according to the following instructions:

  • The access to the FTP site should be performed with FileZilla


  • Protocol: FTP

  • Host (FTP sever): tremplin.epfl.ch
.
  • Username: <USER>@mmspgdata.epfl.ch

  • Password: <PASSWORD>
  • FTP port: 21


NOTE: The username includes “@mmspgdata.epfl.ch”


The USER/PASSWORD will be sent to you by email.