MICAS: Multi-grained In-Context Adaptive Sampling for 3D Point Cloud Processing

To complement the main paper, this supplementary material provides additional details and insights, structured as follows: • Sec. A details the implementation of MICAS. • Sec. B introduces the model variants of MICAS. • Sec. C presents an additional ablation study on the impact of the number of candidate prompts in query-specific prompt sampling. • Sec. D provides an ablation study to further demonstrate the robustness of our proposed MICAS. • Sec. E offers additional qualitative analysis by visualizing sampled points. • Sec. F discusses the limitations of our approach and its broader impacts. A. Implementation Details Following PIC [2], we sample 1, 024 points from each point cloud and segment them into 64 patches, each containing 32 neighboring points. PointNet [5] is used as the task encoder, point encoder, and prompt sampling module. For task-adaptive point sampling, we set the initial learning rate to 1e -4, reducing it to 1e -6 over 60 epochs using a Cosine Annealing Scheduler [4], with a batch size of 72 and a sampling loss hyperparameter α of 0.5. For query-specific prompt sampling, 8 candidate prompts are randomly selected per query, with a learning rate of 1e -5, decay to 1e -6, 30 training epochs, and a batch size of 9.