Closed YodaEmbedding closed 1 year ago
Are you going to update the result image by adding elic to the graph? Thank you.
This is only the paper's reported results -- not the CompressAI implementation. The datasets and chosen GMM $K=3$ may be different from a potential future implementation from CompressAI.
But for reference, here's a plot anyways:
Generated via:
python -m compressai.utils.plot \
-o kodak-psnr.png \
--title="Performance evaluation on Kodak - PSNR (RGB)" \
--axes 0.00 2.25 26 41 \
-f \
results/image/kodak/paper-elic2022_mse.json \
results/image/kodak/vtm.json \
results/image/kodak/paper-cheng2020-checkerboard-6M_mse.json \
results/image/kodak/compressai-cheng2020-anchor_mse_cuda.json \
results/image/kodak/compressai-mbt2018_mse_cuda.json \
results/image/kodak/av1.json \
results/image/kodak/compressai-mbt2018-mean_mse_cuda.json \
results/image/kodak/bpg_444_x265_ycbcr.json \
results/image/kodak/hm.json \
results/image/kodak/compressai-bmshj2018-hyperprior_mse_cuda.json \
results/image/kodak/compressai-bmshj2018-factorized_mse_cuda.json \
results/image/kodak/webp.json \
results/image/kodak/jpeg2000.json \
results/image/kodak/jpeg.json
Side note: I wonder if it would have been more consistent naming to call elic2022 one of the following:
Is ELIC the entire model or just the entropy/context model? What would we call it if we applied ELIC's "uneven channel groups + checkerboard" context model to a different set of $g_a, g_s, h_a, h_s$? Questions!
Official results from Checkerboard and ELIC papers.
Measured using WebPlotDigitizer.