Backward Compatible Object Detection Using HDR Image Content

Convolution Neural Network (CNN)-based object detection models have achieved unprecedented accuracy in challenging detection tasks. However, existing detection models (detection heads) trained on 8-bits/pixel/channel low dynamic range (LDR) images are unable to detect relevant objects under lighting conditions where a portion of the image is either under-exposed or over-exposed. Although this issue can be addressed by introducing High Dynamic Range (HDR) content and training existing detection heads on HDR content, there are several major challenges, such as the lack of real-life annotated HDR dataset(s) and extensive computational resources required for training and the hyper-parameter search. In this paper, we introduce an alternative backwards-compatible methodology to detect objects in challenging lighting conditions using existing CNN-based detection heads. This approach facilitates the use of HDR imaging without the immediate need for creating annotated HDR datasets and the associated expensive retraining procedure. The proposed approach uses HDR imaging to capture relevant details in high contrast scenarios. Subsequently, the scene dynamic range and wider colour gamut are compressed using HDR to LDR mapping techniques such that the salient highlight, shadow, and chroma details are preserved. The mapped LDR image can then be used by existing pre-trained models to extract relevant features required to detect objects in both the under-exposed and over-exposed regions of a scene. In addition, we also conduct an evaluation to study the feasibility of using existing HDR to LDR mapping techniques with existing detection heads trained on standard detection datasets such as PASCAL VOC and MSCOCO. Results show that the images obtained from the mapping techniques are suitable for object detection, and some of them can significantly outperform traditional LDR images.

Abstract licence: CC BY-NC-ND

56 more references, whose wording the paper’s licence keeps private.

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.