Grasp Detection under Occlusions Using SIFT Features

Table 1

Performance of different methods on Cornell Grasp Dataset.

Approach	Algorithm	Accuracy (%)
		IW	OW
Jiang et al. [36]	Fast search	60.5	58.3
Lenz et al. [21]	SAE, struct.	73.9	75.6
Redmon and Angelova [8]	AlexNet, MultiGrasp	88.0	87.1
Guo et al. [22]	ZF-net, hybrid network	93.2	89.1
Chu et al. [1]	ResNet-50 FCGN	97.7	94.9
Li et al. [19]	Key point-based scheme	96.05	96.5
This paper	The proposed algorithm	97.2	92.5

Bold values indicate the performance of our algorithm on Cornell Grasp Dataset. IW: image-wise. The dataset is divided based on image randomly. Each image has an equal probability to be trained or tested. This is a common way to test the generalization of the network to new orientation and position about objects it has seen before. OW: object-wise. The dataset is divided based on object instances. Objects in training set and test set can be different. OW is used to test the generalization ability of a network about new object.