In order to find the physical parameters which determine the accuracy of photometric redshifts, we compare the spectroscopic and photometric redshifts (photo-z's) for a large sample of ~ 80000 SDSS–2MASS galaxies. Photo-z's in this paper are estimated by using the artificial neural network photometric redshift method (ANNz). For a subset of ~40000 randomly selected galaxies, we find that the photometric redshift recovers the spectroscopic redshift distribution very well with rms of 0.016. Our main results are as follows: (1) Using magnitudes directly as input parameters produces more accurate photo-z's than using colors; (2) The inclusion of 2MASS (J, H, Ks) bands does not improve photo-z's significantly, which indicates that near infrared data might not be important for the low-redshift sample; (3) Adding the concentration index (essentially the steepness of the galaxy brightness profile) as an extra input can improve the photo-z's estimation up to ~ 10 percent; (4) Dividing the sample into early- and late-type galaxies by using the concentration index, normal and abnormal galaxies by using the emission line flux ratios, and red and blue galaxies by using color index (g – r), we can improve the accuracy of photo-z's significantly; (5) Our analysis shows that the outliers (where there is a big difference between the spectroscopic and photometric redshifts) are mainly correlated with galaxy types, e.g., most outliers are late-type (blue) galaxies.