Efficient Co-Approximate Parallel Compressive Depth Reconstruction on FPGA
Date:
This presentation introduced a co-approximate framework for parallel compressive depth reconstruction on FPGA.
Key highlights included:
- Single Photon Time-of-Flight (ToF) sensing with eye safety and SNR considerations
- Precision and power approximation combining arithmetic approximation, voltage undervolting, and frequency overscaling
- Parallel processing design with up to 64 processing units
- Approximate accelerator generation workflow with Xilinx HLS-based synthesis
- Cost, performance, and fidelity trade-offs under co-approximation
- Demonstrated up to 60% power savings, significant speedups, and maintained depth image fidelity
The talk concluded with perspectives on error modeling, mixed-precision strategies, and compact low-power LiDAR sensing.
