Background
During my computer-vision algorithm internship at VIA Technologies, I worked on safety-related detection using company factory and automotive data. These are my contributions within the company setting, rather than the complete team project.
Goal
Improve detection of small and occluded objects and integrate inference into an embedded safety-monitoring workflow.
My contributions
- Reproduced and fine-tuned YOLOv8, SAM and Florence-v2.
- Applied augmentation and confidence fusion to improve detection.
- Optimized inference using PyTorch/TensorRT and fast pre-screening.
- Integrated outputs to support rapid local updates on company data.
Approach
- Reproduced and fine-tuned models on factory and automotive data, comparing single-model baselines.
- Applied augmentation and confidence fusion to small and occluded objects.
- Introduced fast pre-screening and TensorRT optimization into inference.
Results
- Resume-reported accuracy increased from 60% to 85%.
- Resume-reported mAP improvement was approximately 12% over single-model baselines; relative versus absolute improvement remains unspecified.
- Resume-reported throughput increased from 3 FPS to 10 FPS.
These are workflow results reported in the resume. Sample size, accuracy definition, mAP thresholds, the meaning of the 12% improvement, hardware and team attribution remain to be clarified; no additional claims are made.