The canopy of trees plays an important role in the ecological system of forest. Its cover, distribution, structure are highly relevant to the function of water cycling, carbon storage, and climate modulating of forest. At individual level, accurate tree crown masks are the bases to acquire precise locations, distribution, and structural parameters of canopy. Therefore, accurate individual tree crown (ITC) segmentation has become a key topic of forestry that supports elaborate forest monitoring, biodiversity assessment, and ecological analyses. With the rapid development remote sensing and easy accessibility of the high-resolution earth observation data, fine-grained canopy observation at individual tree level has been feasible in practice. And, deep learning technologies have achieved impressive performances on the tasks of instance segmentation which promote the accuracy of ITC delineation dramatically. This research aims to fully explore the performance of the SOTA instance segmentation networks, i.e., accuracy, generalization, and transferability, on the task of ITC segmentation. Especially, the performance of the large model, e.g., Segment Anything Model (SAM), is estimated as well. Comprehensive datasets for ITC segmentation with considerate quality, quantity, and diversity is adopted for network training and testing. Multiple ITC segmentation methods are developed by training the SOTA instance segmentation networks by datasets. The precision of the ITC segmentation method is evaluated based on standardized metrics. And, the generalization and transferability are estimated by comparing the segmentation results from testing sets that contains data from various forest types and scenarios. The method with the best performance is the network with HTC baseline and CB-ResNet50 backbone that trained by early-stop scheme, and its AP50 and AP75 achieves 40.98% and 21.25%, respectively.