Semantic Segmentation on MFNet-IR (val)
50.78mIoUUNIV(LoRA)
Evaluation results
| Method | Links | |
|---|---|---|
| UNIV(LoRA)Training Paradigm=Pre-training with extra params, Pre-train Dataset=IN1K+MSIP, Pre-train Method=PCCL, Epoch=100, Total Params(M)=88.7, Training Params(M)=1.82025.09 | 50.78 | |
| MCMAETraining Paradigm=Cross-domain fine-tuning, Pre-train Dataset=IN1K, Pre-train Method=MIM, Epoch=1600, Total Params(M)=88.72025.09 | 50.29 | |
| PADTraining Paradigm=Pre-training with extra params, Pre-train Dataset=IN1K+MSIP, Pre-train Method=MIM, Epoch=100, Total Params(M)=87.0, Training Params(M)=1.22025.09 | 46.89 | |
| MILANTraining Paradigm=Cross-domain fine-tuning, Pre-train Dataset=IN1K, Pre-train Method=MIM, Epoch=400, Total Params(M)=85.82025.09 | 45.08 | |
| MAETraining Paradigm=Cross-domain fine-tuning, Pre-train Dataset=IN1K, Pre-train Method=MIM, Epoch=1600, Total Params(M)=85.82025.09 | 44.47 | |
| MAETraining Paradigm=Full pre-training from IN1K, Pre-train Dataset=IN1K+MSIP, Pre-train Method=MIM, Epoch=1600, Total Params(M)=85.8, Training Params(M)=85.82025.09 | 43.97 | |
| MoCo v3Training Paradigm=Cross-domain fine-tuning, Pre-train Dataset=IN1K, Pre-train Method=CL, Epoch=300, Total Params(M)=85.82025.09 | 43.23 | |
| MAETraining Paradigm=Full pre-training from scratch, Pre-train Dataset=MSIP, Pre-train Method=MIM, Epoch=400, Total Params(M)=85.8, Training Params(M)=85.82025.09 | 39.58 |