Exploring Budgeted Image Classification with Content-Sensitive Resource Allocation

This paper formulates Budgeted Image Classification as allocating a batch of images across multiple decision points with different computational costs, maximizing accuracy under a changing compute budget. The resulting resource-allocation integer program is NP-hard. The authors introduce a continuous relaxation that yields a content-agnostic strategy, then propose a content-sensitive strategy that uses image-specific information. According to the abstract, experiments show that content-sensitive allocation performs better. The paper also derives conditions for suitable decision points and examines failure cases, identifying directions for future work.