How yizhou-h/trajectory-mamba-cvpr works
Main Components
a. Data Ingestion
• Argoverse1 and Argoverse2 datasets (images + annotations.json)
b. Preprocessing & Encoding
• Data loader modules
• Joint polyline encoding strategy
c. Core Model
• Encoder (selective SSM‐based attention layers)
• Decoder (trajectory generation head)
d. Configuration & Checkpoint Management
• configs/main_train.yml
• checkpoints/argoverseX_ckpt.pth
e. Evaluation & Output
• eval.py script
• GPU execution environment (CUDA_VISIBLE_DEVICES)
Relationships and Data Flow
• Data Ingestion → Preprocessing & Encoding
(polyline construction + feature embedding)
• Preprocessed Features → Encoder
(Selective State‐Space Model replaces quadratic self‐attention)
• Encoder Outputs → Decoder
(forecast future trajectories)
• Decoder Outputs → Evaluation Script
(compute metrics, export results)
• Config Files → Drive parameters for all pipeline stages
• Checkpoints → Model weight initialization before inference/evaluation
Architectural Patterns
• Encoder–Decoder Architecture
• Self‐Attention redesigned via Linear Selective SSM
• Joint Polyline Encoding for multi‐agent context
• Modular design (data, model, config, evaluation)
Call Out Critical Technologies & Frameworks
• PyTorch (model implementation)
• CUDA/NVIDIA GPUs (inference & training)
• Anaconda environment + environment.yaml
• YAML‐based configuration management
• Argoverse public datasets
– Clear component labels (e.g., “SelectiveSSMAttention”, “PolylineEncoder”)