Table 5.
Without | With | ||||||||
---|---|---|---|---|---|---|---|---|---|
OTS | FT | 1channel | large | OTS | FT | 1channel | large | ||
Without | OTS | — | 0.65 | 0.74 | 0.73 | 0.46 | 0.38 | 0.40 | 0.59 |
FT | 0.65 | — | 0.81 | 0.80 | 0.38 | 0.42 | 0.43 | 0.64 | |
1channel | 0.74 | 0.81 | — | 0.93 | 0.41 | 0.43 | 0.47 | 0.71 | |
large | 0.73 | 0.80 | 0.93 | — | 0.40 | 0.43 | 0.47 | 0.71 | |
With | OTS | 0.46 | 0.38 | 0.41 | 0.40 | — | 0.32 | 0.33 | 0.39 |
FT | 0.38 | 0.42 | 0.43 | 0.43 | 0.32 | — | 0.35 | 0.42 | |
1channel | 0.40 | 0.43 | 0.47 | 0.47 | 0.33 | 0.35 | — | 0.45 | |
large | 0.59 | 0.64 | 0.71 | 0.71 | 0.39 | 0.42 | 0.45 | — |
Our experiments are grouped into three categories. First, “Without” and “With” non-image features. Second, transfer-learning with off-the-shelf (OTS) and fine-tuned (FT) models. Third, from scratch where “1channel” refers to same input size as in transfer-learning but changed number of channels. “large” means we changed the input dimensions to 448 × 448 × 1. We identify three clusters: all models under “With”, models trained from scratch and “Without”, and the “OTS” model.