Benchmarking deep models on salient object detection

Abstract

The performance discrepancies caused by different implementation details may obscure the actual progress of the Salient Object Detection (SOD) task. In this paper, we construct a SALient Object Detection (SALOD) benchmark to ensure a fair and comprehensive evaluation by unifying the implementation details of SOD methods. By doing so, we can reveal the reasons behind recent progress by analyzing the impact of network structure and optimization strategy. Based on the experimental results, we first find that U-shaped networks, both older and more recent variants, achieve better performance than other structures. Second, optimization strategy, e.g., training strategy and loss function, significantly impacts SOD accuracy. Finally, we provide a new perspective to validate the generalizability of SOD methods on objectness shifting. Code is available at https://github.com/moothes/SALOD.