Which output to trust? None of them. Atlases provide a segmentation for a particular subject and indicate only a rough estimate of where structures might be in another subject.
Part of your question deals with templates, and I'll start with that. The N27 brain is based on a single individual that was scanned 27 times while the 152 and 452 templates are based on datasets from the MNI group who created these templates from average of aligned datasets of 152 and 452 datasets, respectively. These three templates were converted from their original template spaces of MNI and MNI_ANAT to the Talairach space using either a manual or the Brett transform (a two-step affine transformation) procedure.
Regarding the atlases, the TT_daemon is based on the Talairach-Tourneaux atlas of a post-mortem analysis of half of an individual brain. The original procedure for MRI data assumes a manual procedure will allow structures in the brain to correspond to the atlas. Under the manual procedure, only the AC-PC line is forced to correspond with only the general size of the brain made to fit the template. In contrast, the auto_tlrc script aligns the data to a template using a single affine transformation using a least-squares minimization between the dataset and the template. The Talairach Daemon provides an atlas, but not a directly comparable template.
Ideally, one would want a template very similar to the subjects in the study and an atlas based on that same template; however, this situation is rarely available. One can still use a template similar to their subjects and transform the template to the space of an available atlas or align to one of the group templates mentioned above. Still the atlas will not provide an exact nor even a close to exact location or extent of structures in a particular subject. One must instead use the atlas only as rough guide for where regions are located. The shape and location of regions and the method of alignment to the template space will have a strong effect on the overlap of the regions with a functional cluster. So overlap and exact position may be off by several millimeters from a structure in a dataset and the corresponding structure in an atlas. The whereami program and the interactive results in the AFNI GUI show the locations for many voxels in a neighborhood around a voxel for exactly this reason. We provide several atlases because they all provide different kinds of regions, and those regions are generated from templates in different ways. All are useful yet may disagree with each other.
Another way of looking at the same issue of structural variability is by using a probabilistic atlas. We provide the Eickhoff-Zilles probabilistic atlases with AFNI and will soon provide several probabilistic atlases from Rutvik Desai's group. With these atlases, for any voxel location, the probability is computed for finding a particular structure from the group of subjects upon which the atlas is based. The alignment procedures and atlas region definition method will have an effect on the results.
Usually you will want to use the atlases that provide the regions in which you are interested. How you use these atlases is another question. Many new features regarding templates and atlases will soon be available as part of the AFNI package, and you will be able to add new atlases quickly and easily to AFNI. Stay tuned. In the meantime, you can take a look at the current atlases in the abin directly as datasets and this presentation from one of our recent classes:
[
afni.nimh.nih.gov]
The output from the overlap mask result of whereami shows the sub-brick of the atlas dataset where the structure is found. The code is the intensity value for a structure. Both are not strictly necessary from a user perspective, but you can use this information to look at the atlas datasets directly.