Incorporating to object detection

[ogencoglu]

Thanks for the interesting work.

Can ensLoss be utlized in object detection tasks in combination with traditional detection losses somehow? Or is it only for classification?

[statmlben]

Thank you for the excellent suggestions and for your encouragement! I'm actually considering extending the ensloss idea to more general problems. The main logic behind this problem is that for a given evaluation metric, there are many valid losses. For instance, in binary classification, SVM (hinge loss), logistic loss, and boosting exponential loss all aim to maximize accuracy. In more general cases, there are indeed many losses that can be optimized, but most lack a clear relationship with the evaluation metric. Ensembling in such cases may offer some practical benefits, but the underlying principle seems somewhat questionable.

I genuinely want to extend this idea to general problems. To do so, we need to identify the conditions under which a loss is valid for a given evaluation metric, and then design ensloss based on these conditions. However, most problems are not as straightforward as classification in deriving these conditions. Nevertheless, we are actively working to make progress on this front.

Let's keep this issue here, and I hope that I can address it:)