抄録
CH-004
Activity recognition based on egocentric object detection
Saptarshi Sinha・Hiroki Ohashi・Mitsuhiro Okada・Takuto Sato・Katsuyuki Nakamura(所属なし)
This paper studies first person activity recognition using egocentric videos. Object information is important for activity recognition but how much fine-grained object knowledge is helpful for the process is still an unresolved research area. We conducted an extensive study on how much fine-grained object location is useful for classifying activities. We experimentally found that too much fine or course grained information can harm activity recognition. We concatenated the object features from fine-tuned YOLOv3 to the features calculated by pretrained 3D-ResNext for activity recognition. We achieved a f-measure of 0.669, which is higher than when only 3D-ResNext features are used(0.653).