"Fast and Efficient Transformer-based Method for Bird's Eye View Instance Prediction"

Bird's Eye View (BEV) Lean transformer for Instance Prediction slashes self-driving prediction costs while matching accuracy.

https://arxiv.org/abs/2411.06851

Original Problem 🎯:

Current self-driving vehicle systems separate detection, tracking, and prediction stages, leading to accumulated errors. Existing solutions have high processing times and parameter counts, making real-world deployment challenging.

-----

Solution in this Paper 🛠️:

→ Introduces a Bird's Eye View (BEV) instance prediction architecture focusing only on instance segmentation and flow prediction

→ Uses EfficientNet-B4 for multi-camera feature extraction and BEV projection

→ Implements two parallel SegFormer-based branches for segmentation and flow prediction

→ Features a flow warping mechanism to track instances across frames

→ Offers two configurations: full version (13M parameters) and tiny version (7.42M parameters)

-----

Key Insights 💡:

→ Simplified paradigm focusing on just two tasks (segmentation and flow) can match SOTA performance

→ Efficient transformer architecture can significantly reduce parameters while maintaining accuracy

→ Flow warping at individual BEV position level minimizes instance association errors

-----

Results 📊:

→ Achieves 53.7 VPQ (Video Panoptic Quality) at short ranges, outperforming PowerBEV (53.4)

→ Reduces parameters from 39.13M (PowerBEV) to 13.46M (full) and 7.42M (tiny)

→ Decreases latency to 60-63ms compared to PowerBEV (70ms)