running model: fully connected to fuse views in a ring, average voting (no attention), square ring setting:
class_score = test_score
class_score = class_score.reshape(-1,N_RING,20)
arg_mean = np.mean(class_score,1)
preds = np.argmax(arg_mean,1)
-
install xvbf (x virtual frame buffer) (if use a headless linux instance)
-
trimesh
-
tf & pytorch.
app.py: main app server
classìfy model: input 26 imgs, output class scores.
extract_feature: extract resnet-50 2048-d for each imgs.
render_26: render 26 colored view around the objects.