chanyn
commented
5 years ago You can edit files of training and testing for COCO and VOC following VG and ADE, we have provided corresponding gt graph of these two datasets. Same settings are used for all datasets in our paper.
if args.dataset == "vg": args.imdb_name = "vg_train" args.imdbval_name = "vg_val" args.set_cfgs = ['ANCHOR_SCALES', '[2, 4, 8, 16, 32]', 'MAX_NUM_GT_BOXES', '50'] cls_r_prob = pickle.load(open('data/graph/vg_graph_r.pkl', 'rb')) cls_r_prob = np.float32(cls_r_prob) cls_a_prob = pickle.load(open('data/graph/vg_graph_a.pkl', 'rb')) cls_a_prob = np.float32(cls_a_prob) elif args.dataset == "ade": args.imdb_name = "ade_train_5" args.imdbval_name = "ade_val_5" args.set_cfgs = ['ANCHOR_SCALES', '[2, 4, 8, 16, 32]', 'MAX_NUM_GT_BOXES', '50'] cls_r_prob = pickle.load(open('data/graph/ade_graph_r.pkl', 'rb')) cls_r_prob = np.float32(cls_r_prob) cls_a_prob = pickle.load(open('data/graph/ade_graph_a.pkl', 'rb')) cls_a_prob = np.float32(cls_a_prob) elif args.dataset == "vgbig": args.imdb_name = "vg_train_big" args.imdbval_name = "vg_val_big" args.set_cfgs = ['ANCHOR_SCALES', '[2, 4, 8, 16, 32]', 'MAX_NUM_GT_BOXES', '50'] cls_r_prob = pickle.load(open('data/graph/vg_big_graph_r.pkl', 'rb')) cls_r_prob = np.float32(cls_r_prob) cls_a_prob = pickle.load(open('data/graph/vg_big_graph_a.pkl', 'rb')) cls_a_prob = np.float32(cls_a_prob)
看这里并没有写上coco和pascal voc的数据集