Happy to share that our "Contrastive Pretraining for Visual Concept Explanations of Socioeconomic Outcomes" paper (https://t.co/dOFvGZjwSL) was accepted at the #EarthVision workshop at #CVPR2024.
with @Alevering1, Lars Penning, Dario Oliveira @diego_marcos_g, and @xiaoxiang_zhu
@cvml_mpiinf@kerstingAIML We now move to our spotlight talks. The first speaker, Ivica Obadic, talks about "Recent Trends, Challenges, and Limitations of Explainable AI in Remote Sensing".
Further, analyzing the model's conceptual sensitivity can shed light on new insights into urban studies. For example, increasing the amount of vegetation in low-income areas improves the model's perception of their income score.
This enables the model to associate the concepts learned with the #TCAV approach with continuous intervals of socioeconomic outcomes. For example, dense residential areas resemble low liveability while more liveable areas are those close to natural areas and vegetation.
Our approach can be used to interpret how #urban concepts relate to #socioeconomic outcomes in computer vision models.
The contrastive pretraining with Rank-N-Contrast helps to increasingly order the instances in the latent space according to their socioeconomic outcomes
Happy to share that our "Contrastive Pretraining for Visual Concept Explanations of Socioeconomic Outcomes" paper (https://t.co/dOFvGZjwSL) was accepted at the #EarthVision workshop at #CVPR2024.
with @Alevering1, Lars Penning, Dario Oliveira @diego_marcos_g, and @xiaoxiang_zhu