This work aims to develop a robustness mobile-friendly model with small model size, high inference speed and reasonable memory usage for the portrait segmentation. We introduced high-resolution features in the encoder part to supply more information and utilized two segmentation head, one of it served as refinement stage to improve accuracy across all categories. To leverage datasets with different annotation, we adopt an unified training strategy with adaptive loss design for different data. We lowered memory usage by reducing the embedding dimension and substitute concatenate operator with series sum inside segmentation head without sacrificing accuracy.

We introduce a novel and unified WB algorithm that could handle both raw device RGB (scene referred raw image) and standard RGB images (output referred non-raw image). Our algorithm is semantic aware that we put emphasis on skin tone and are easily to extend to other objects colors. Our method is easy to personalize and more convenient to integrate into products with good performance compared to large deep learning models.

We deisgn a network for image retrieval without human annotations like bounding box, which utilizes coarse-scale feature fusion and generates the attentive local features via combining the information from different intermediate layers. Two attentions blocks help extracted detailed information.

We propose an end-to-end fashion analysis network that addresses category classification and attribute prediction simultaneously, via improving the resolution of heatmaps through upsampling for more accurate landmark localization. A novel attention mechanism is introduced: Landmark heatmaps are used as references to generate a unified attention, so that the network has enough information to enhance or reduce features.