Zero-shot Quantization for Object Detection
Keywords: Zero-shot quantization, Object detection, Synthetic data, Fine-tuning efficiency, Feature distillation
TL;DR: Introducing ZSQ framework for object detection, utilizing synthetic calibration sets for privacy and efficiency. Enhances mAP accuracy and training speed over traditional methods.
Abstract: Zero-shot quantization (ZSQ) has achieved remarkable success in classification tasks by leveraging synthetic data for network quantization without accessing the original training data. However, when applied to object detection networks, current ZSQ methods fail due to the inherent complexity of the task, which encompasses both localization and classification challenges. On the one hand, the precise location and size of objects within the samples for object detection remain unknown and elusive in zero-shot scenarios, precluding artificial reconstruction without ground-truth information. On the other hand, object detection datasets typically exhibit category imbalance, and random category sampling methods designed for classification tasks cannot capture this information.
To tackle these challenges, we propose a novel ZSQ framework specifically tailored for object detection. The proposed framework comprises two key steps: First, we employ a novel bounding box and category sampling strategy in the calibration set generation process to infer the original training data from a pre-trained detection network and reconstruct the location, size and category distribution of objects within the data without any prior knowledge. Second, we incorporate feature-level alignment into the Quantization Aware Training (QAT) process, further amplifying its efficacy through the integration of feature-level distillation.
Extensive experiments conducted on the MS-COCO and Pascal VOC datasets demonstrate the efficiency and state-of-the-art performance of our method in low-bit-width quantization. For instance, when quantizing YOLOv5-m to 5-bit, we achieve a 4.2\% improvement in the mAP metric, utilizing only about 1/60 of the calibration data required by commonly used LSQ trained with full trainset.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4776
Loading