Table 2. NAS results: the italiced row... | Emerald Publishing

Table 2.

NAS results: the italiced row indicates the best-performing model, while the italicized row represents the baseline model

	Coarse filtering				Fine filtering
Detection model	Memory (GiB)	Frame rate (FPS)	Accuracy (mAP@0.5:0.95)	Evaluation	Best hyperparameters	Best training configuration	Best accuracy (Link to mail.mAP@0.5:0.95)
CRPN	11.2	23	0.399	Pruned because of high memory usage
Conditional DETR	9.1	30	0.411	Passed	Backbone: Resnet50	Epoch: 150 Batch size: 16	0.445
DAB-DETR	9.1	26	0.423	Passed	Backbone: Resnet50	Epoch: 50Batch size: 16	0.456
DETR	7.2	34	0.399	Passed	Backbone: Resnet50	Epoch: 150 Batch size: 16	0.433
DINO	11.1	9	0.490	Pruned because of high memory usage and low inference speed
EfficientNet	11.1	5	0.404	Pruned because of high memory usage and low inference speed
FreeAnchor	4.9	30	0.387	Passed	Backbone: ResNeXt 101	Epoch: 300 Batch size: 128	0.419
RetinaNet	5.0	19	0.365	Pruned because of low inference speed
RTMDet	8.9	16	0.446	Pruned because of low inference speed
SSD	9.9	30	0.255	Pruned because of low accuracy
YOLOv5	4.7	29	0.338	Pruned because of low accuracy
YOLOX	7.6	24	0.405	Passed	Backbone: CSPDarknet	Epoch: 150 Batch size: 64	0.440

	Coarse filtering				Fine filtering
Detection model	Memory (GiB)	Frame rate (FPS)	Accuracy (mAP@0.5:0.95)	Evaluation	Best hyperparameters	Best training configuration	Best accuracy (mAP@0.5:0.95)
CRPN	11.2	23	0.399	Pruned because of high memory usage
Conditional DETR	9.1	30	0.411	Passed	Backbone: Resnet50	Epoch: 150 Batch size: 16	0.445
DAB-DETR	9.1	26	0.423	Passed	Backbone: Resnet50	Epoch: 50Batch size: 16	0.456
DETR	7.2	34	0.399	Passed	Backbone: Resnet50	Epoch: 150 Batch size: 16	0.433
DINO	11.1	9	0.490	Pruned because of high memory usage and low inference speed
EfficientNet	11.1	5	0.404	Pruned because of high memory usage and low inference speed
FreeAnchor	4.9	30	0.387	Passed	Backbone: ResNeXt 101	Epoch: 300 Batch size: 128	0.419
RetinaNet	5.0	19	0.365	Pruned because of low inference speed
RTMDet	8.9	16	0.446	Pruned because of low inference speed
SSD	9.9	30	0.255	Pruned because of low accuracy
YOLOv5	4.7	29	0.338	Pruned because of low accuracy
YOLOX	7.6	24	0.405	Passed	Backbone: CSPDarknet	Epoch: 150 Batch size: 64	0.440

Tracking model	Coarse filtering				Fine filtering
Tracking model	Accuracy (MOTA)		Occlusion handling (IDSwR)	Evaluation	Best neural architecture		MOTA
ByteTrack	0.781		0.102	Passed	Detector: DAB-DETR		0.786
DeepSORT	0.620		0.159	Pruned because of low occlusion handling
OC-SORT	0.778		0.041	Passed	Detector: DAB-DETR		0.781
Tracktor	0.641		0.061	Passed	ReID: Resnet50 Detector: Faster-RCNN		0.693

Tracking model	Coarse filtering				Fine filtering
Tracking model	Accuracy (MOTA)		Occlusion handling (IDSwR)	Evaluation	Best neural architecture		MOTA
ByteTrack	0.781		0.102	Passed	Detector: DAB-DETR		0.786
DeepSORT	0.620		0.159	Pruned because of low occlusion handling
OC-SORT	0.778		0.041	Passed	Detector: DAB-DETR		0.781
Tracktor	0.641		0.061	Passed	ReID: Resnet50 Detector: Faster-RCNN		0.693

Source(s): Created by authors