Multibiometric recognition has become one of the most important solutions for enhancing overall personal recognition performance due to several inherent limitations of unimodal biometrics, such as nonuniversality and unacceptable reliability. However, most existing multibiometrics fuse completely different biometric traits based on addition schemes, which usually require several sensors and make the final feature sets large. In this paper, we propose a joint multi-instance hand-based biometric feature learning method for biometric recognition. Specifically, we first exploit the important direction data from multi-instance biometric images. Then, we simultaneously learn the discriminative features of multi-instance biometric traits and exploit the collaborative representations of multi-instance biometric features such that the final joint multi-instance feature descriptor is compact. Moreover, the importance weights of different biometric instances can be adaptively learned. Experimental results on the baseline multi-instance finger-knuckle-print and palmprint databases demonstrate the promising effectiveness of the proposed method.