add project proposal

.github/prompts/2.project_proposal.prompt.md

@@ -0,0 +1,21 @@
+# Task on Project Proposal
+
+## Summary
+
+We have already done our pilot study report, and now we have to write a project proposal.
+
+for details of project, refer to `1.main_task.prompt.md` file.
+
+for already-done pilot study, refer to `main.tex` file.
+
+To complete this task you should compose a short post outlining your research question
+and hypothesis and comment on the posts of two other students with related work.
+
+The project proposal should include the following sections:
+
+- Research Question and Hypothesis
+- Literature Review
+- Methodology
+- Planning
+
+Generate a testable research hypothesis based on your research question. This should be something that hasn’t already been tested extensively and has the potential to generate valuable information in your field of study. If the research generates multiple hypotheses that can be easily tested together (for example testing a number of different algorithms for the same problem), it is okay for your study to include these hypotheses.

assignment_1/project_proposal.md

@@ -0,0 +1,29 @@
+# Project Proposal: Attention-Enhanced 3D Point Cloud Segmentation for Railway Foreign Object Detection
+
+## Research Question and Hypothesis
+
+**Research Question:** Can attention-based 3D point cloud segmentation architectures significantly improve the detection of rare foreign objects on railway tracks compared to standard RandLA-Net?
+
+**Hypothesis:** A hybrid architecture combining Attention mechanisms with RandLA-Net's efficient sampling will achieve at least 65% IoU for foreign object detection, while maintaining computational efficiency suitable for real-time monitoring.
+
+## Literature Review
+
+Recent advancements in transformer-based architectures like Point Transformer V3 [Wu et al., 2024] have shown superior performance in capturing fine-grained features. Meanwhile, approaches like Sphere Transformer [Lai et al., 2023] specifically address LiDAR's unique distribution patterns. The class imbalance problem has been partially addressed by RAPiD-Seg [Li et al., 2024] through range-aware feature embeddings.
+
+## Methodology
+
+We will develop a hybrid architecture integrating:
+
+1. RandLA-Net's efficient random sampling for computational feasibility
+2. Point Transformer's vector attention mechanisms for capturing fine details
+3. Class-balanced loss functions to address the extreme imbalance (0.001% foreign objects)
+4. Data augmentation techniques including selective jittering [Park et al., 2024]
+
+Evaluation will use precision, recall, and IoU metrics, with special emphasis on foreign object performance. Computational efficiency will be benchmarked against RandLA-Net baseline.
+
+## Planning
+
+- Weeks 1-2: Architecture design and implementation
+- Weeks 3-4: Training and hyperparameter optimization
+- Weeks 5-6: Evaluation and comparison with baseline
+- Weeks 7-8: Documentation and reporting