Mobile crowd photography (MCP) is a widely used technique in crowd sensing. In MCP, a picture stream is generated when delivering intermittently to the backend server by participants. Pictures contributed later in the stream may be semantically or visually relevant to previous ones, which can result in data redundancy. To meet diverse constraints (e.g., spatiotemporal contexts, single or multiple shooting angles) on the data to be collected in MCP tasks, a data selection process is needed to eliminate data redundancy and reduce network overhead. This issue has little been investigated in existing studies. To address this requirement, we propose a generic data collection framework called PicPick. It first presents a multifaceted task model that allows for varied MCP task specification. A pyramid tree (PTree) method is further proposed to select an optimal set of pictures from picture streams based on multi-dimensional constraints. Experimental results on two real-world datasets indicate that PTree can effectively reduce data redundancy while maintaining the coverage requests, and the overall framework is flexible