update ignore

.gitignore

@@ -0,0 +1,11 @@
+.idea
+__pycache__
+.DS_Store
+.vscode
+csrc/eval_cylib/*.so
+logs/
+.ipynb_checkpoints
+dist_mats
+submit
+third_party
+*.out

config/defaults.py

@@ -95,7 +95,7 @@
 # List of the dataset names for testing
 _C.DATASETS.TEST_NAMES = "market1501"
 _C.DATASETS.ROOT_DIR = ('/Volumes/Data/比赛/行人重识别2019/data')
-_C.DATASETS.PRELOAD_IMAGE = True   # 是否将图像加载到内存
+_C.DATASETS.PRELOAD_IMAGE = True    # 是否将图像加载到内存
 
 # -----------------------------------------------------------------------------
 # DataLoader
@@ -173,4 +173,4 @@
 # ---------------------------------------------------------------------------- #
 # Misc options
 # ---------------------------------------------------------------------------- #
-_C.OUTPUT_DIR = "/Volumes/Data/比赛/行人重识别2019/logs/"
+_C.OUTPUT_DIR = "/Volumes/Data/比赛/行人重识别2019/logs/"

configs/softmax_triplet.yml

@@ -32,7 +32,7 @@ SOLVER:
   EVAL_PERIOD: 10
 
 TEST:
-  IMS_PER_BATCH: 64
+  IMS_PER_BATCH: 8
   WEIGHT: "path"
 
 

configs/softmax_triplet_resnext.yml

@@ -31,7 +31,7 @@ SOLVER:
   EVAL_PERIOD: 10
 
 TEST:
-  IMS_PER_BATCH: 32
+  IMS_PER_BATCH: 16
   WEIGHT: "path"
 
 

data/datasets/competition1910.py

@@ -11,6 +11,10 @@
 
 from .bases import ImageDataset
 
+def argsort(seq):
+    # http://stackoverflow.com/questions/3071415/efficient-method-to-calculate-the-rank-vector-of-a-list-in-python
+    return sorted(range(len(seq)), key=seq.__getitem__)
+
 
 class Competition1910(ImageDataset):
     """
@@ -34,8 +38,8 @@ def __init__(self, root, test_phase=False, **kwargs):
             self.query_dir = osp.join(self.dataset_dir, 'myval_query')
             self.gallery_dir = osp.join(self.dataset_dir, 'myval_gallery')
         else:
-            #self.train_dir = osp.join(self.dataset_dir, 'mytrain')
-            self.train_dir = osp.join(root, self.test_data_dir, 'myquery') # not used
+            self.train_dir = osp.join(self.dataset_dir, 'mytrain')    # not used
+            #self.train_dir = osp.join(root, self.test_data_dir, 'mytrain') # not used
             self.query_dir = osp.join(root, self.test_data_dir, 'myquery')
             self.gallery_dir = osp.join(root, self.test_data_dir, 'mygallery')
 
@@ -70,9 +74,10 @@ def _process_dir(self, dir_path, relabel=False):
         #print('dir_path', dir_path, 'img_paths', len(img_paths), img_paths[0])
         #pattern = re.compile(r'([-\d]+)_c(\d)')
 
+
         DEBUG = False
         if DEBUG:
-            img_paths = img_paths[:200]
+            img_paths = img_paths[:1000]
 
         pid_container = set()
         for img_path in img_paths:
@@ -84,10 +89,13 @@ def _process_dir(self, dir_path, relabel=False):
         pid2label = {pid: label for label, pid in enumerate(pid_container)}
 
         dataset = []
+
+        original_filenames = []
         for img_path in img_paths:
             #pid, camid = map(int, pattern.search(img_path).groups())
             pid = int(img_path.split('/')[-1].split('_')[0])
             camid = int(img_path.split('/')[-1].split('_')[1][:-4])
+            original_filenames.append(img_path.split('/')[-1].split('_')[1])
 
             if pid == -1:
                 continue  # junk images are just ignored
@@ -98,4 +106,10 @@ def _process_dir(self, dir_path, relabel=False):
             if relabel: pid = pid2label[pid]
             dataset.append((img_path, pid, camid))
 
+        # sort dataset items according to filename in ascending order
+        idx = argsort(original_filenames)
+        dataset = [dataset[i] for i in idx]
+
+        #print('dataset', dataset[:10])
+
         return dataset

engine/inference.py

@@ -11,7 +11,7 @@
 from data.datasets.eval_reid import evaluate
 from data.prefetcher import data_prefetcher
 from utils.re_ranking import re_ranking, compute_distmat_using_gpu
-from utils.distance import low_memory_local_dist
+from utils.distance import low_memory_local_dist, fast_local_dist
 import h5py
 import time
 
@@ -340,47 +340,37 @@ def compute_distmat(cfg, num_query, feats, feats_flipped, local_feats, local_fea
     local_distmat1 = None
     local_distmat2 = None
 
+    aligned = True
+
     if use_rerank:
+        #print('len(local_feats)', len(local_feats))
         if len(local_feats) > 0 and use_local_feature:
             # if True:
             # calculate the local distance
             lqf = lqf.permute(0, 2, 1)
             lgf = lgf.permute(0, 2, 1)
-            local_qg_distmat = low_memory_local_dist(lqf.numpy(), lgf.numpy(), aligned=True)
-            local_qq_distmat = low_memory_local_dist(lqf.numpy(), lqf.numpy(), aligned=True)
-            local_gg_distmat = low_memory_local_dist(lgf.numpy(), lgf.numpy(), aligned=True)
-            local_distmat1 = np.concatenate(
-                [np.concatenate([local_qq_distmat, local_qg_distmat], axis=1),
-                 np.concatenate([local_qg_distmat.T, local_gg_distmat], axis=1)],
-                axis=0)
-            del local_qg_distmat, local_qq_distmat, local_gg_distmat
+            local_distmat1  = fast_local_dist(lqf.numpy(), lgf.numpy(), aligned=aligned)
+            del lqf, lgf
 
         query_num = qf.size(0)
         gallery_num = gf.size(0)
 
-        distmat = compute_distmat_using_gpu(qf, gf)
-        distmat1 = re_ranking(distmat, query_num, gallery_num, k1=6, k2=2, lambda_value=0.3,
-                              local_distmat=local_distmat1,
-                              theta_value=theta,
+        distmat = compute_distmat_using_gpu(qf, gf, local_distmat1, theta=theta)
+        distmat1 = re_ranking(distmat, query_num, gallery_num, k1=12, k2=2, lambda_value=0.3,
+                              #local_distmat=local_distmat1,
+                              #theta_value=theta,
                               only_local=False)
         del distmat, local_distmat1
 
         if len(local_feats) > 0 and use_local_feature:
             # flipped
             lqf = lqf_flipped.permute(0, 2, 1)
             lgf = lgf_flipped.permute(0, 2, 1)
-            local_qg_distmat = low_memory_local_dist(lqf.numpy(), lgf.numpy(), aligned=True)
-            local_qq_distmat = low_memory_local_dist(lqf.numpy(), lqf.numpy(), aligned=True)
-            local_gg_distmat = low_memory_local_dist(lgf.numpy(), lgf.numpy(), aligned=True)
-            local_distmat2 = np.concatenate(
-                [np.concatenate([local_qq_distmat, local_qg_distmat], axis=1),
-                 np.concatenate([local_qg_distmat.T, local_gg_distmat], axis=1)],
-                axis=0)
-            del local_qg_distmat, local_qq_distmat, local_gg_distmat
-
-        distmat = compute_distmat_using_gpu(qf_flipped, gf_flipped)
-        distmat2 = re_ranking(distmat, query_num, gallery_num, k1=6, k2=2, lambda_value=0.3, local_distmat=local_distmat2, theta_value=theta,
-                         only_local=False)  # (current best)
+            local_distmat2 = fast_local_dist(lqf.numpy(), lgf.numpy(), aligned=aligned)
+            del lqf, lgf, lqf_flipped, lgf_flipped
+
+        distmat = compute_distmat_using_gpu(qf_flipped, gf_flipped, local_distmat2, theta=theta)
+        distmat2 = re_ranking(distmat, query_num, gallery_num, k1=12, k2=2, lambda_value=0.3, only_local=False)  # (current best)
         del distmat, local_distmat2
         distmat = (distmat1 + distmat2) / 2
 
@@ -435,11 +425,19 @@ def inference_aligned_flipped(
             if len(ret) == 4:
                 gf, bn_gf, lf, bn_lf = ret
                 gff, bn_gff, lff, bn_lff = ret_flip
-            else:
+            elif len(ret) == 2:
                 gf, bn_gf = ret
                 gff, bn_gff = ret_flip
                 lf, bn_lf = None, None
                 lff, bn_lff = None, None
+            elif ret is not tuple:
+                gf = bn_gf = ret
+                gff = bn_gff = ret_flip
+                lf, bn_lf = None, None
+                lff, bn_lff = None, None
+            else:
+                # print('ret', ret.size())
+                raise Exception("Unknown model returns, length = ", len(ret))
 
         # 4 features
         gfs.append(gf.cpu())
@@ -471,30 +469,27 @@ def inference_aligned_flipped(
     g_pids = np.asarray(pids[num_query:])
     g_camids = np.asarray(camids[num_query:])
 
-    logger.info(f"use_cross_feature = {use_cross_feature}, use_local_feature = {use_local_feature}, use_rerank = {use_rerank}")
+    logger.info(f"use_local_feature = {use_local_feature}, use_rerank = {use_rerank}")
 
-    if use_cross_feature:
-        logger.info("Computing distmat with bn_gf (+ lf)")
-        distmat2 = compute_distmat(cfg, num_query, bn_gfs, bn_gfs_flipped, lfs, lfs_flipped, theta=0.45,
+
+    logger.info("Computing distmat with bn_gf")
+    distmat2 = compute_distmat(cfg, num_query, bn_gfs, bn_gfs_flipped, lfs, lfs_flipped, theta=0.45,
                                    use_local_feature=use_local_feature, use_rerank=use_rerank)
-        distmat = distmat2
-        #distmat = (distmat1 + distmat2) / 2
-    else:
-        logger.info("Computing distmat with gf + bn_lf")
-        distmat1 = compute_distmat(cfg, num_query, gfs, gfs_flipped, bn_lfs, bn_lfs_flipped, theta=0.95,
+
+
+    logger.info("Computing distmat with gf + bn_lf")
+    distmat1 = compute_distmat(cfg, num_query, gfs, gfs_flipped, bn_lfs, bn_lfs_flipped, theta=0.95,
                                    use_local_feature=use_local_feature, use_rerank=use_rerank)
-        distmat = distmat1
-        #distmat1 = None
-        #distmat2 = None
 
-    #distmat = original_distmat
-    #distmat[:, new_gallery_index] = distmat1 - 100
+    for theta in np.linspace(0, 1, 21):
+        #theta = 0.55
+        distmat = distmat1 * (1 - theta) + distmat2 * theta
 
-    cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
-    logger.info(f"mAP: {mAP:.1%}")
-    for r in [1, 5, 10]:
-        logger.info(f"CMC curve, Rank-{r:<3}:{cmc[r - 1]:.1%}")
-    logger.info(f"Score: {(mAP + cmc[0]) / 2.:.1%}")
+        cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
+        logger.info(f"theta: {theta:.2%} mAP: {mAP:.1%}")
+        for r in [1, 5, 10]:
+            logger.info(f"CMC curve, Rank-{r:<3}:{cmc[r - 1]:.1%}")
+        logger.info(f"Score: {(mAP + cmc[0]) / 2.:.1%}")