SCUT-Ego-Finger Dataset is for research on egocentric vision hand detection and fingertip detection. It includes 93729 frames from 24 videos captured and labeled by 24 volunteers. Detail scenes include: Avenue (4058), Basketball Field (2894), Chinese Book (3001), Classroom (4868), Computer Screen (5088), East Canteen (3738), East Lake (4151), Football Field (4145), Fountain (4158), Lcy-Lab (4084), Liwu Building (5488), NO.1 North Dorm (4281), NO3. North Dorm (4679), North Canteen (4314), North Lake (3419), North Library (2495), Renwen Building (3611), Supermarket (1682), Teaching Building (4368), Tennis Field (5124), West Canteen (4185), Wyx-Lab (3222), Yifu Building 1st Floor (3806), Yifu Building 2nd Floor (2870).
The dataset is collected for solving problem of ego-vision fingertip detection and hand detection, which may conduct challenges such as background complexity, Illumination change, hand shape and hand color diversity, motion blur, so on and forth. Therefore the dataset deliberately covers different type of frames. It is feasible to applied Deep Learning on the dataset. For further evaluation of the dataset, please visit https://github.com/hyichao/Performer, or contact email@example.com
DeepFinger: A Cascade Convolutional Neuron Network Approach to Finger Key Point Detection in Egocentric Vision with Mobile Camera
Published in IEEE Conference on System, Man and Cybernatic, 2015. Available at IEEE Xplore: DeepFinger: A Cascade Convolutional Neuron Network Approach to Finger Key Point Detection in Egocentric Vision with Mobile Camera
We released SCUT-HCII Ego-Finger Dataset in September, 2015.
The dataset is organized with one folder and a related txt file. The folder stores images of the set under a certain scene, like Fountain. Images inside are named as “I_Fountain_xxx.png”, where alphabet “I” represents gesture type, while “Fountain” tells the collection location and “xxx” is simple the image number. As for txt file, we arrange it as following section b), containing file name, 4 float labels of bounding box top-left and bottom-right coordinates (x, y), and then followed with 4 float labels of finger key point coordinate (x, y). More details are revealed in b).
a) Image samples
I_Fountain_xxx.png tlx tly brx bry ftx fty jntx jnty tjntx tjnty 0 0 0 0
I_RenwenBuilding_xxx.png tlx tly brx bry ftx fty jntx jnty tjntx tjnty 0 0 0 0
I_Chinesebook_xxx.png tlx tly brx bry ftx fty jntx jnty tjntx tjnty 0 0 0 0
tlx && tly: bounding box top-left point (x, y)
brx&& bry: bounding box bottom-right point (x, y)
ftx && fty: fingertip point (x, y)
jntx&&jnty: index finger joint point (x, y)
tjntx&&tjnty: tail finger joint point (x, y)
The number stored is normalized with image width and image height,
i.e. I_Chinesebook_xxx.png 0.4 0.4 0.6 0.6 0.5 0.5 0.55 0.55 0.55 0.55 0 0 0 0
While the image is 640*480, the actual coordinate of points are
(256, 192), (384,288), (320,240), (352,264) , (352,264)
Please fill in the APPLICATION FORM, follow the note inside and send an email
If the dataset link is not available due to network problem or server error, please contact firstname.lastname@example.org for other approaches.