How sanvisharma850/intentformer-trajectory-prediction works
IntentFormer is a notebook-based research workflow for predicting pedestrian and cyclist trajectories from nuScenes. The documented path runs from dataset preprocessing to model training and evaluation; the model combines temporal, social, and LiDAR context to produce multimodal trajectories and intent labels. The graph preserves the README’s documented model components, losses, and metrics. Source excerpts were not available, so notebook internals and some handoffs are not verified; those links are dashed where inferred.
Main entry points: 02-model-train-v2.ipynb Model training · 01-dataprep-v2.ipynb Data preprocessing · 03-eval-v2.ipynb Model evaluation · evaluation_summary.md Evaluation results
Components 19
Data preparation
- Data preprocessing
01-dataprep-v2.ipynb - Trajectories
01-dataprep-v2.ipynb - Neighbor data
01-dataprep-v2.ipynb - LiDAR features
01-dataprep-v2.ipynb - Normalization stats
01-dataprep-v2.ipynb
Prediction model
- Temporal encoder
02-model-train-v2.ipynb - Social encoder
02-model-train-v2.ipynb - LiDAR features
02-model-train-v2.ipynb - Trajectory modes
02-model-train-v2.ipynb - Intent classifier
02-model-train-v2.ipynb - GMM predictions
02-model-train-v2.ipynb
Training
- Model training
02-model-train-v2.ipynb - GMM-NLL loss
02-model-train-v2.ipynb - Diversity loss
02-model-train-v2.ipynb - Intent CE loss
02-model-train-v2.ipynb
Evaluation
- Model evaluation
03-eval-v2.ipynb - Prediction metrics
03-eval-v2.ipynb - Evaluation results
Training_Output/evaluation_summary.md
Other components
- nuScenes dataset
Connections 25
- nuScenes dataset → Data preprocessing: provides data
README.md - Data preprocessing → Trajectories: produces
README.md - Data preprocessing → Neighbor data: produces
README.md - Data preprocessing → LiDAR features: produces
README.md - Data preprocessing → Normalization stats: produces
README.md - Temporal encoder → Trajectory modes: encodes history
README.md - Social encoder → Trajectory modes: provides context
README.md - LiDAR features → Trajectory modes: provides features
README.md - Trajectory modes → GMM predictions: predicts modes
README.md - Intent classifier → GMM predictions: classifies intent
README.md - Model training → GMM-NLL loss: optimizes with
README.md - Model training → Diversity loss: optimizes with
README.md - Model training → Intent CE loss: optimizes with
README.md - Model training → Temporal encoder: trains model
README.md - Model training → Social encoder: trains model
README.md - Model training → LiDAR features: trains model
README.md - Model training → Trajectory modes: trains model
README.md - Model training → Intent classifier: trains model
README.md - Trajectories → Model training: feeds training
- Neighbor data → Model training: feeds training
- LiDAR features → Model training: feeds training
- Normalization stats → Model training: feeds training
- Model evaluation → Prediction metrics: computes
README.md - Model evaluation → Evaluation results: records results
- GMM predictions → Model evaluation: evaluates predictions