Conclusion

To further enhance the developed MVP, I have summarized what aspects should be considered:

1. Multi-task Capability:
   • Expanding the model’s capabilities to effectively handle multiple tasks.
   • Specifically, utilizing BEV feature for lane and drivable area detection, and leveraging PV feature map for traffic light detection.
2. Performance Evaluation, Metrics
   • The need for metrics tailored to automotive applications rather than relying solely on metrics from the computer vision domain.
   • Beyond the conventional Average Precision (AP) metric, there’s a strong emphasis on metrics that incorporate temporal awareness (similar to tracking) for downstream tasks.
   • Furthermore, the metrics must consider the object’s location, as the impact of False Positives or False Negatives at close distances greatly differs from that at longer distances.
3. Matching Function
   • Personally, I hold the view that neither Intersection-over-Union (IoU) metrics (Waymo and KITTI) nor center distance metrics (nuScenes) are the optimal choices for performance evaluation.
   • In scenarios like auto parking, even a minor localization error of 10 cm could lead to a significant accident, underscoring the importance of accurate object contour detection.
   • However, in highway driving scenarios, a longitudinal localization error of 1 m at an object distance of 100 m is negligible, while detecting the relative speed and lateral position of the object are more important.
4. Real-time Processing:
   • Ensuring that the model can efficiently run on embedded PC hardware.
   • Exploring techniques such as TensorRT and INT8 quantization to optimize inference speed.

By addressing these considerations, we can refine the MVP and advance toward the development of a more sophisticated and capable product.