Lightweight Object Detection Ensemble Framework for Autonomous Vehicles in Challenging Weather Conditions

Table 3

Comparison of previous models to the proposed model for OD performance on DAWN dataset [82, 83].

Models	Training dataset	Voting strategy	mAP (of rain and fog) (in %)
Previous models (mAP is for all 4 categories)
Faster R-CNN with several RPNs (with Faster R-CNN backbone) [84]	DAWN	n/a	89.48
Previous models
Standard data augmentation [82]	Pretrained on ImageNet and fine-tuned on the” clear” split of BDD100k-cls	n/a	23.3
AMDA [82]	Pretrained on ImageNet and fine-tuned on the” clear” split of BDD100k-cls	n/a	25.55
Ensemble (AMDA, AMDA) [82]	Pretrained on ImageNet and fine-tuned on the” clear” split of BDD100k-cls	n/a	25.8
RoHL (AMDATV-ftGauss, AMDA-ftCont) [82]	Pretrained on ImageNet and fine-tuned on the HF and LF biases.	n/a	24.9
Proposed models
RetinaResnet50	COCO	Affirmative	32.75