Rate-incremental learning (high to low) on Kodak VTM 22.0
0.86BD-Rate (%) (Old Bitstreams)Our model w/ KR
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Our model w/ KRStrategy=Knowledge Replay (KR)2024.02 | 0.86 | 4.26 | 2.56 | |
| GMA-VR w/ KRStrategy=Knowledge Replay (KR)2024.02 | 1.75 | 6.92 | 4.34 | |
| Our modelStrategy=pre-trained2024.02 | 1.87 | — | — | |
| GMA-VRStrategy=pre-trained2024.02 | 4.43 | — | — | |
| GMA-VR w/ FT Enc. & Dec.Strategy=Fine-tuning Encoder & Decoder2024.02 | 10 | 9.11 | 9.56 | |
| Our model w/ FT Enc. & Dec.Strategy=Fine-tuning Encoder & Decoder2024.02 | 14.33 | 5.18 | 9.76 | |
| MSH-VR w/ KRStrategy=Knowledge Replay (KR)2024.02 | 18.9 | 29.9 | 24.4 | |
| MSH-VRStrategy=pre-trained2024.02 | 26.7 | — | — | |
| MSH-VR w/ FT Enc. & Dec.Strategy=Fine-tuning Encoder & Decoder2024.02 | 35.1 | 37.6 | 36.4 |