Core Concepts
HalluciDet proposes a novel approach using privileged information from pre-trained detectors to guide image-to-image translation for object detection.
Abstract
The paper introduces HalluciDet, an IR-RGB image translation model focusing on object detection. It leverages privileged information from pre-trained RGB detectors to enhance detection accuracy in the IR modality. The proposed method aims to reduce the domain gap between IR and RGB modalities by generating meaningful representations for accurate detections. Experimental results on LLVIP and FLIR datasets demonstrate significant improvements over traditional image translation techniques and fine-tuning methods.
Directory:
Introduction
Hardware sensors advance data collection for deep learning algorithms.
Related Work
Object detection methods categorized as two-stage and one-stage detectors.
Proposed Method
HalluciDet framework utilizing privileged information for image-to-image translation.
Experimental Results and Analysis
Evaluation on LLVIP and FLIR datasets showcasing improved performance over traditional methods.
Conclusion
HalluciDet offers a framework using privileged information to enhance object detection through image translation.
Stats
モデルはFaster R-CNNを使用し、AP@50で88.34の検出精度を達成。
FLIRデータセットでは、Fine-tuningに比べて70%のトレーニングサンプルで性能向上を示す。
HalluciDetはResNet34バックボーンを使用し、FLIRデータセットでAP@50が71.58となる。