What Time Tells Us? An Explorative Study of Time Awareness Learned from Static Images, Dongheng Lin^*, Han Hu^*, Jianbo Jiao^†

• A new paradigm to learn visual cues for timestamps leading to time-aware understanding on random visual inputs.
• Built a new benchmark dataset with 130k samples paired with accurate time/geolocation metadata.
• Tested the learned representations on various downstream tasks (Image Retrieval, Video Scene Recognition and Time-conditioned Image Editing).

Audio-Visual Separation with Hierarchical Fusion and Representation Alignment, Han Hu^*, Dongheng Lin^*, Qiming Huang, Yuqi Hou, Hyung Jin Chang, Jianbo Jiao^†

• We proposed Hierarchical Fusion Strategy that integrates middle and late fusion to enhance separation across diverse sound characteristics.
• We introduced a new loss term for the task that aligns U-Net’s latent audio features with pre-trained audio embeddings (CLAP) to improve semantic quality and reduce modality gap.

SPSS: A Salience-based Poisoning Selection Strategy for Selecting Backdoor Attack Victims, Zihan Lyu¹, Dongheng Lin¹, Shengkai Sun², Jie Zhang^†, Ruiming Zhang²

• Designed an algorithm uses Salience Metric to evaluate sample feature significance towards backdoor learning process.
• We managed to realize a more data-efficient backdoor attack to DNN models achieved the same attack success rate to vanilla backdoor attack with only 38.44% of poisoned samples.

• FUJICHROME Velvia 100, taken at my hometown.