Requirements Addressed {#did_TIADALG motion segmentation and object detection requirements}
Introduction
Purpose
Short Application Description
The models include semantic segmentation, motion segmentation and a joint model that does both together. The models are trained in PyTorch and were finally converted to ONNX format. Cityscapes dataset was used for the training and the accuracy numbers reported are on this dataset.
Input and Output format
Pre-Processing for input Image
- Step 1: Resize to 1024x512
- Step 2: Divide by 255.0
- Step 3: Mean subtract ([0.485, 0.456, 0.406])
- Step 4: Multiply ([1/0.229, 1/0.224, 1/0.225])
Semantic Segmentation
- Input: RGB Image preprocessed as dscussed above
- Output: integer output plane, indicating the class id of segmentation.
- Validation Accuracy: 70.4 mIoU
Motion Segementation
- Input1: RGB Image preprocessed as dscussed above
- Input2: 3 planes of DOF data
- Output : integer output plane, indicating the class id of segmentation.
- Validation Accuracy: 85.0 mIoU
Directory Structure
Diagrams
Sequence Diagram
Component Interaction
OpenVX Graph
Resource usage
Error handling
Interface
Design Analysis and Resolution (DAR)
Design Decision : none
na
Design Criteria: none
na
Design Alternative: none
na
Design Alternative: none
na
Final Decision
na