Good Teacher Makes Good Student: A Discriminative-Aware Knowledge Preservation Approach for Zero-Shot Sketch-Based Image Retrieval

Sketch-Based Image Retrieval (SBIR) is widely used in animation, e-commerce, and security. In these real-world applications, the classes of retrieval may be very different from the training classes, making it a zero-shot SBIR problem. Most methods in the literature resort to bridging the semantic gap between the sketch and image domains by learning a common space with a pre-trained model on a large dataset as the base network, and then finetuning on the target SBIR datasets. In this process, the acquired knowledge of the pre-trained model may be lost, resulting in performance degradation. To tackle this problem, we propose a teacher-student network architecture, which consists of a teacher network using the pre-trained model and a student network whose output is guided by the teacher network. Instead of introducing supplementary semantics in the teacher network, we adopt a more powerful pre-trained model as the teacher network and further enhance its discriminative capability by adding a hard-coded margin based on the prediction probability. The student network is then fine-tuned by using the teacher network's output as the learning target. Experiments on two benchmark datasets show that the proposed approach outperforms the state-of-the-art methods by more than 10%, which verifies that the prior knowledge can be better preserved by a good teacher network, which can make the student network good too.