X-ray image guidance enables percutaneous alternatives to complex procedures.\nUnfortunately, the indirect view onto the anatomy in addition to projective\nsimplification substantially increase the task-load for the surgeon. Additional\n3D information such as knowledge of anatomical landmarks can benefit surgical\ndecision making in complicated scenarios. Automatic detection of these\nlandmarks in transmission imaging is challenging since image-domain features\ncharacteristic to a certain landmark change substantially depending on the\nviewing direction. Consequently and to the best of our knowledge, the above\nproblem has not yet been addressed. In this work, we present a method to\nautomatically detect anatomical landmarks in X-ray images independent of the\nviewing direction. To this end, a sequential prediction framework based on\nconvolutional layers is trained on synthetically generated data of the pelvic\nanatomy to predict 23 landmarks in single X-ray images. View independence is\ncontingent on training conditions and, here, is achieved on a spherical segment\ncovering (120 x 90) degrees in LAO/RAO and CRAN/CAUD, respectively, centered\naround AP. On synthetic data, the proposed approach achieves a mean prediction\nerror of 5.6 +- 4.5 mm. We demonstrate that the proposed network is immediately\napplicable to clinically acquired data of the pelvis. In particular, we show\nthat our intra-operative landmark detection together with pre-operative CT\nenables X-ray pose estimation which, ultimately, benefits initialization of\nimage-based 2D/3D registration.\n