This repo is focus on multi-labbel scene classification and gives 2 methods. The dataset is from here.
This solution is mainly from here, and I managed to modify something like optimizer. But Essentially, this solution transformed multi-label classification into binary classification.
| Accuracy | Value |
|---|---|
| Hamming Loss | 0.139 |
| Subset Accuracy(Exact Match) | 0.52375 |
In this solution, the trained model resnet50 from keras is used as image feature extractor. Then these features are inputed to multi-label algorithms(like Classifer Chains, Random k-Labelsets, etc.)The main reason why I choose resnet50 is because the output dimension of other models is too high (like VGG16 or VGG19 has 1*7*7*512). The output dimension of resnet50 is just 1*1*1*2048.
| Accuracy | Binary Relevance | Calibrated Label Ranking | Random k-Labelsets | MLKNN |
|---|---|---|---|---|
| Hamming Loss | 0.065500 | 0.067000 | 0.058000 | 0.066500 |
| Subset Accuracy | 0.720000 | 0.717500 | 0.762500 | 0.740000 |