L-TReiD: Logic Tensor Transformer for Re-identification

Alessandro Sebastian Russo, Francesco Manigrasso, Fabrizio Lamberti, Lia Morra

Abstract: This article proposes a Neuro-Symbolic (NeSy) machine learning approach to Object Re-identification. NeSy is an emerging branch of artificial intelligence which combines symbolic reasoning and logic-based knowledge representation with the learning capabilities of neural networks. Since object re-identification involves assigning the identity of the same object across different images and different conditions, such a task could benefit greatly from leveraging the logic capabilities of a NeSy framework to inject prior knowledge about invariant properties of the objects. To test this assertion, we combined the Logic Tensor Networks (LTNs) NeSy framework with a state-of-the-art Transformer-based Re-Identification and Damage Detection Network (TransRe3ID). The LTN incorporates prior knowledge about the properties that two instances of the same object have in common. Experimental results on the Bent&Broken Bicycle re-identification dataset demonstrate the potential of LTNs to improve re-identification systems and provide novel opportunities to identify pitfalls during training.