ACTNET: End-to-End Learning of Feature Activations and Multi-stream Aggregation for Effective Instance Image Retrieval

We propose a novel CNN architecture called ACTNET for robust instance image retrieval from large-scale datasets. Our key innovation is learnable activation layer designed to improve the signal-to-noise ratio of deep convolutional feature maps. Further, we introduce controlled multi-stream aggregation, where complementary features different layers are optimally transformed and balanced using our layers, before aggregation into global descriptor. Importantly, parameters blocks explicitly trained, together with parameters, in an end-to-end manner minimising triplet loss. This means that network jointly learns filters their optimal tasks. To knowledge, this first time parametric functions have been used control learn aggregation. conduct in-depth experimental study on three non-linear functions: Sine-Hyperbolic, Exponential modified Weibull, showing while all bring significant gains Weibull function performs best thanks its ability equalise strong activations. The results clearly demonstrate significantly enhances discriminative power features, improving over state-of-the-art