Retraining decoders and hypercoders on encoder-generated quantized latents gives consistent BD-rate savings of 1 to 2 percent for scalar and trellis-coded quantization.
Testing on the higher resolution TecNick set produced BD-rates of −1.98% and −2.29% as listed in Table 4
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Optimizing Learned Image Compression on Scalar and Entropy-Constraint Quantization
Retraining decoders and hypercoders on encoder-generated quantized latents gives consistent BD-rate savings of 1 to 2 percent for scalar and trellis-coded quantization.