Research map / Architectures and efficiency / Quantization
QuartDepth: Post-Training Quantization for Real-Time Depth Estimation on the Edge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QuartDepth, a 4-bit post-training quantization framework (LogNP activation polishing, activation-loss weight compensation, Hessian-based AdaRound weight reconstruction) plus a programmable ASIC accelerator for monocular depth estimation models.
Problem
Foundational monocular depth estimation models are too computationally and memory heavy to deploy on resource-limited edge devices and ASICs, and standard PTQ at W4A4/W4A8 degrades accuracy due to per-channel activation outliers in depth decoders.
Approach
Smooth decoder activation outliers per channel with a log-based polishing transform (LogNP) before activation quantization, compensate activation quantization error by updating weights layer-wise, then quantize weights with a KFAC-Fisher-guided AdaRound reconstruction, and co-design a fused-kernel programmable ASIC to hide the polishing overhead.
Key results
- On Metric3D ViT-Large under W4A4 on NYUv2, QuartDepth reaches AbsRel 0.133 and delta1 0.815 versus 0.217 / 0.598 for BRECQ, while W4A8 is nearly lossless (AbsRel 0.071 vs 0.067 float32).
- On the 28nm ASIC, Metric3D ViT-Small at 256 resolution under W4A4 reaches 38.3 ms latency (26.11 FPS), a 2.0x speedup over Float32 (76.8 ms).
- Silicon area is 29.22 mm2 for Float32, 23.94 mm2 for W4A8 and 24.35 mm2 for W4A4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QuartDepth W4A4 instead of BrecQ W4A4 | Metric3D ViT-Large · NYUv2 | AbsRel: 0.469 -> 0.097 | Helped |
| Use QuartDepth W4A4 instead of BrecQ W4A4 | Metric3D ViT-Large · KITTI | AbsRel: 0.434 -> 0.070 | Helped |
| Use QuartDepth W4A8 instead of BrecQ W4A8 | Metric3D ViT-Large · NYUv2 | AbsRel: 0.076 -> 0.071 | Helped |
| Use QuartDepth W4A8 instead of BrecQ W4A8 | Metric3D ViT-Large · KITTI | AbsRel: 0.057 -> 0.055 | Helped |
| Use QuartDepth W4A4 instead of BrecQ W4A4 | Depth Anything ViT-Large · NYUv2 | AbsRel: 0.342 -> 0.070 | Helped |
| Use QuartDepth W4A4 instead of BrecQ W4A4 | Depth Anything ViT-Large · KITTI | AbsRel: 0.385 -> 0.059 | Helped |
| Use QuartDepth W4A8 instead of BrecQ W4A8 | Depth Anything ViT-Large · SUN RGB-D | AbsRel: 0.489 -> 0.394 | Helped |
| Use QuartDepth W4A8 instead of BrecQ W4A8 | Depth Anything ViT-Large · KITTI | AbsRel: 0.051 -> 0.060 | Hurt |
| Use QuartDepth W4A8 instead of BrecQ W4A8 | Depth Anything ViT-Large · vKITTI2 | AbsRel: 0.606 -> 0.108 | Hurt |
| Use W4A4 instead of Float32 | Metric3D ViT-Small on ASIC at 256 resolution · ASIC inference | latency: 76.8 (1↖) -> 38.3 (2.0↖) | Helped |
Surprising findings
- Naive per-channel PTQ methods such as minmax and EMA collapse at W4A4 on Metric3D (e.g., minmax AbsRel 1.265 on NYUv2 with ViT-Small), while the proposed method stays usable.
Limitations
- Evaluated only on Metric3D and Depth Anything MDE models; generalization to other architectures is not shown.
- Hardware results come from RTL synthesis and a performance emulator rather than a fabricated chip.
- Relies on a calibration set of 32 images and a KFAC-approximated Hessian, which may be sensitive to calibration choices.
Experimental setup
Datasets: NYUv2 KITTI
Benchmarks: NYUv2 KITTI SUN RGB-D iBims-1 HyperSim vKITTI2 DIODE Outdoor
Metrics: AbsRel delta1 delta2 delta3 RMSE Silog latency FPS power efficiency (GMAC/W)
Baselines: OBS minmax EMA percentile AdaRound BrecQ Float32
Builds on: AdaRound BRECQ Metric3D Depth Anything KFAC
Closest papers by meaning
- Predicting High-precision Depth on Low-Precision Devices Using 2D Hilbert Curves (ICML 2025 · similarity 0.86)
- DCDepth: Progressive Monocular Depth Estimation in Discrete Cosine Domain (NeurIPS 2024 · similarity 0.85)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.85)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.85)
- BetterDepth: Plug-and-Play Diffusion Refiner for Zero-Shot Monocular Depth Estimation (NeurIPS 2024 · similarity 0.85)
- Efficient Depth Estimation for Unstable Stereo Camera Systems on AR Glasses (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.