CLIC: http://www.compression.cc/challenge/

linux

• Prepare dataset,(e.g. KODAK), put it under the *./dataset* directory

• run bash jpeg_compression.sh KODAK, bash jpeg2000_compression.sh KODAK, bash webp_compression.sh KODAK, to compress KODAK dataset with *jpeg*, *jpeg2000* and *webp* codec seprately.

• Before we use

  hevc

  and

  hevc-jem

  codec, we have to transefer original image file into YUV file.

  • cd ./utils/yuvtools

  • matlab -nodesktop -nodisplay
    

    • MATLAB command:
    • rgb_dir='../../dataset/KODAK';yuv_dir='../../datast/KODAK_YUV';ext='png'
    • run scrirgb2yuv, all png file in rgbdir will be save into yuvdir as .yuv file

• cd ../.., run bash hevc_compress.sh ./dataset/KODAK_YUV ./compressed_set/KODAK/hevc ./compressed_set/KODAK/hevc_result *and* bash hevc-jem_compress.sh ./dataset/KODAK_YUV ./compressed_set/KODAK/jem-hevc ./compressed_set/KODAK/jem-hevc_result, to get *hevc* and *hevc-jem* compressed stream and results.

• With previous steps, we get jpeg, jpeg2000, webp, hevc and hevc-jem compressed streams/results

• For

  tfrnn codec

  , change directory into ./

  Train/tfnn

  compression/image_encoder, then run

  python compress.py --ori_dir=(path to your image dataset directory) --cps_dir=(path to save the reconstructed images)
  

  • since the stream size compressed by tfrnn codec is revelant to the iteration, so we dont have to save the compressed stream to calculate the stream size.

• run python image_metric.py --ref_dir='./dataset/KODAK' --cps_dir='./compressed_set/KODAK/webp, metric will be saved into *KODAK_webp.csv* in the current folder.
• after you get all the KODAK_(codec).csv, move all the csv files into
  • python plot_scatter.py --dataset='KODAK' --ref_image_dir='./dataset/KODAK' --metric_dir='./' --metric_name='Y-psnr' --metric_index=8 --output_dir='./metric_data/KODAK-YPSNR' --debug=True, metric plot will be save into ./metric_data/KODAK-YPSNR
• To calculate BDBR and BDPSNR, run python calc_bdbr_bdpsnr.py --dataset=KODAK --ref_image_dir='./dataset/KODAK' --metric_dir='./' --ref_codec='jpeg', make sure KODAK_(codec).csv is under ./ directory. Codec Jpeg is default reference codec to compare with.

• --input_dir

  • original images directory.

• --codec

  :

  • support jpeg, jpeg2000, webp

• --output_dir

  :

  • All compresed streams will save to this directory.

• --quality

  • between 1 and 100 (worst and best quality)

• DEMO: python compress_image.py --input_dir='/data/sunnycia/image_compression_challenge/dataset/KODAK' --output_dir='/data/sunnycia/image_compression_challenge/compressed_set/KODAK/webp' --quality=50 --codec='webp'

• --yuv_dir

  :

  • original yuv image file directory

• --hm_dir

  :

  • default '_Train/HM-16.9'

• --cfg_path

  :

  • default 'cfg/encoder_lowdelay_main_rext.cfg'

• --qp

  :

  • 0~50, best to worst

• --yuvformat

  :

  • input yuv file format, 420 or 444.

• DEMO: python hevc_compress.py --yuv_dir='/data/sunnycia/image_compression_challenge/dataset/KODAK_YUV' --result_dir='/data/sunnycia/image_compression_challenge/compressed_set/KODAK_YUV' --qp=50

• --ref_dir

  :

  • directory of original images

• --cps_dir

  • directory of compressed images(support jpeg, webp, jp2 stream file)
  • if use for hevc metric extraction, just leave this arg with space '_'

• --color_space

  • default 'YUV'
  • calculate BGR metric or YVU metric

• --output_path

  • Default None, the program will generate the path by itself
  • metric output path

• DEMO: python image_metric.py --ref_dir='/data/sunnycia/image_compression_challenge/dataset/KODAK_PNG' --cps_dir='/data/sunnycia/image_compression_challenge/compressed_set/KODAK/jpeg2000'

• --dataset
  • dataset name in the ./dataset directory, e.g. KODAK
• --refimagedir
  • original image directory
• --metric_dir
  • default ./
  • directory for the csv metric files, e.g
• --metric_name
  • default 'psnr'
  • metric name in plot yaxis.
• --metric_index
  • default 8
  • the col index in the csv file
  • 8-> y-psnr, 9->u-psnr, 10->v-psnr, 11->yuv-psnr
• --output_dir
  • default 'metric_data'
  • output directory of plot
• --debug
  • default False
  • Since the plot lib *plotly* only permit saving 100 figs per day, so set debug=True in case you get a bad plot result.
• DEMO: python plot_scatter.py --dataset='KODAK' --ref_image_dir='/data/sunnycia/image_compression_challenge/dataset/KODAK' --metric_dir='./metric_data' --metric_name='Y-psnr' --metric_index=8 --output_dir='./metric_data/KODAK-YPSNR' --debug=True

• --dataset

  :

  • Dataset name

• --refimagedir

  :

  • original image directory

• --metric_dir

  :

  • default 'jpeg'
  • directory of csv metric file

• --output_dir

  :

  • default './metric_data'
  • bdbr and bdpsnr output file

• --ref_codec

  :

  • default 'jpeg'
  • which codec for reference to calculate bdbr & bdpsnr

• --metric_index

  :

  • default 8
  • psnr metric index

• DEMO:python calc_bdbr_bdpsnr.py --dataset=KODAK --ref_image_dir='/data/sunnycia/image_compression_challenge/dataset/KODAK'