Stereoscopic saliency estimation with background priors based deep reconstruction.

•A deep autoencoder-based reconstruction model is proposed for 3D saliency estimation.•Integrating depth in the initial stage exhibits more decisive results on perception.•Enhancing the learning of background is effective to reduce the false positive rate.•Reconstruction residual is coherent with the cognitive gap to reflect saliency.•Our model shows inspiring results on fixation prediction and proto-object extraction.