This paper presents a new method for small-object recognition from RGB-D data. The method uses a group of trained Gaussian Mixture Models (GMMs) to find SIFT feature matches for object recognition. To save computational time, SIFT features are extracted from a number of isolated image blocks of the image (instead of the entire image). A 2-step approach is employed to produce the image blocks that might contain the target object. First, the saliency of each image pixel is used to extract the salient image blocks. If the target object is found in one of the blocks by the GMM object detector, the method succeeds. Otherwise, the range data is processed to extract the major planar surfaces whose image counterparts are removed from the image to produce isolated image blocks. The GMM-based SIFT matching process is then repeated on these image blocks to detect the target object. Experimental results demonstrate that the proposed method can detect a target object or an object-type with a high reliability. The GMM-based method is scalable and has good parallelism for real-time implementation.