Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

QSCA: Quantization with Self-Compensating Auxiliary for Monocular Depth Estimation

Authors: Jincheol Yang, Jaemin Choi, Matti Zinke, Suk-Ju Kang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that QSCA significantly improves quantized depth estimation performance. On the NYUv2 dataset, it achieves an 11% improvement in ฮด1 accuracy over existing post-training quantization methods. ... We evaluate our framework on NYUv2 [42] and KITTI [43] under two quantization settings: 4/4 and 4/6 (weight/activation bit-widths). Our QSCA consistently outperforms prior PTQ methods... As shown in Tables 1 and 2...
Researcher Affiliation Academia Jincheol Yang Department of Electronic Engineering Sogang University EMAIL Jaemin Choi Department of Electronic Engineering Sogang University EMAIL Matti Zinke Department of Computer Science Sogang University EMAIL Suk-Ju Kang Department of Electronic Engineering Sogang University EMAIL
Pseudocode No The paper describes its methodology using mathematical equations and textual explanations, but it does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes We release our code and checkpoints, including training and evaluation scripts. All datasets used in the paper are open-sourced
Open Datasets Yes We employ five widely used monocular depth estimation benchmarks: NYUv2 [42], KITTI [43], Sintel [45], ETH3D [46], and DIODE [47]. ... We release our code and checkpoints, including training and evaluation scripts. All datasets used in the paper are open-sourced
Dataset Splits No We randomly select 16 samples from each of the NYUv2, and KITTI datasets to calibrate the quantization parameters. ... When finetuning the SCA modules, we perform 1 epoch of training using only a randomly selected 5% subset of the training set. ... For evaluation, we employ five widely used monocular depth estimation benchmarks: NYUv2 [42], KITTI [43], Sintel [45], ETH3D [46], and DIODE [47]. These benchmarks cover diverse domains such as indoor environments, outdoor driving scenes, synthetic imagery, and mixed settings. We select them because they correspond to the official zero shot relative depth evaluation datasets used in the Depth Anything models [7, 8].
Hardware Specification Yes All experiments are conducted with batch size of 1, and a single RTX 4090 GPU.
Software Dependencies No The paper does not explicitly state specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes We randomly select 16 samples from each of the NYUv2, and KITTI datasets to calibrate the quantization parameters. For the calibration strategy, we apply the widely used percentile method [44], employing channel-wise quantization for weights and layer-wise quantization for activations. To initialize the parameters of the SCA modules, we randomly sample a single image from the training set. When finetuning the SCA modules, we perform 1 epoch of training using only a randomly selected 5% subset of the training set. For training, we use the Adam optimizer with a learning rate of 1 ร— 10โˆ’4 and without weight decay. The hyperparameters are set as follows: ฮป1 is set to 1e3, and ฮปfeat is set to 0.5. All experiments are conducted with batch size of 1, and a single RTX 4090 GPU.