store keypoint score for imw2020

README.md

@@ -26,4 +26,10 @@ then call:
 cd /local/aslfeat && python image_matching.py --config configs/matching_eval.yaml
 ```
 
-### 2. Benchmark on FMBench
+### 2. Benchmark on FMBench
+
+Download the data (validation/test) from [here](https://vision.uvic.ca/imw-challenge/index.md), then configure ``configs/imw2020_eval.yaml``, finally call:
+
+```bash
+cd /local/aslfeat && python evaluations.py --config configs/imw2020_eval.yaml
+```

configs/imw2020_eval.yaml

@@ -0,0 +1,20 @@
+data_name: 'imw2020'
+data_split: 'val' # val or test
+data_root: '/local/data/phototourism'
+dump_root: '/local/imw2020_val_aslfeat_2k'
+truncate: [0, null]
+model_path: 'pretrained/aslfeat/model.ckpt-380000'
+overwrite: true
+net:
+  max_dim: 2048
+  config:
+    kpt_n: 2048
+    kpt_refinement: true
+    deform_desc: 1
+    score_thld: 0.5
+    edge_thld: -1
+    multi_level: true
+    nms_size: 3
+    eof_mask: 5
+    need_norm: true
+    use_peakiness: true

datasets/imw2020.py

@@ -63,16 +63,21 @@ def _format_data(self, data):
         seq_folder = os.path.join(self.config['dump_root'], seq_name)
         h5_kpt = os.path.join(seq_folder, 'keypoints.h5')
         h5_desc = os.path.join(seq_folder, 'descriptors.h5')
+        h5_score = os.path.join(seq_folder, 'scores.h5')
 
-        if not os.path.exists(h5_desc) and not os.path.exists(h5_kpt):
+        if not os.path.exists(h5_desc) and not os.path.exists(h5_kpt) and not os.path.exists(h5_score):
             gen_kpt_f = h5py.File(h5_kpt, 'w')
             gen_desc_f = h5py.File(h5_desc, 'w')
+            gen_score_f = h5py.File(h5_score, 'w')
         else:
             gen_kpt_f = h5py.File(h5_kpt, 'a')
             gen_desc_f = h5py.File(h5_desc, 'a')
+            gen_score_f = h5py.File(h5_score, 'a')
 
         if basename not in gen_kpt_f and basename not in gen_desc_f:
             feat = data['dump_data'][0]
             kpt = data['dump_data'][1]
+            score = data['dump_data'][2]
             _ = gen_kpt_f.create_dataset(basename, data=kpt, dtype='f')
             _ = gen_desc_f.create_dataset(basename, data=feat, dtype='f')
+            _ = gen_score_f.create_dataset(basename, data=score, dtype='f')

evaluations.py

@@ -33,9 +33,9 @@ def extract_feat(config):
         try:
             data = next(test_set)
             if config['overwrite'] or not os.path.exists(data['dump_path']):
-                desc, kpt = model.run_test_data(data['image'])
+                desc, kpt, score = model.run_test_data(data['image'])
                 dump_data = {}
-                dump_data['dump_data'] = (desc, kpt)
+                dump_data['dump_data'] = (desc, kpt, score)
                 dump_data['image_path'] = data['image_path']
                 dump_data['dump_path'] = data['dump_path']
                 dataset.format_data(dump_data)

models/cnn_wrapper/aslfeat.py

@@ -128,6 +128,7 @@ def setup(self):
         det_kpt_coord = tf.stack(
             [det_kpt_inds[:, :, 1], det_kpt_inds[:, :, 0]], axis=-1, name='kpt')
         self.layers['kpt'] = det_kpt_coord
+        self.layers['score'] = tf.identity(det_kpt_score, name='score')
 
     def our_score(self, inputs, ksize=3, all_softplus=True, need_norm=True, dilation=1, name='conv'):
         if need_norm:

models/feat_model.py

@@ -0,0 +1,42 @@
+import sys
+
+import cv2
+import numpy as np
+import tensorflow as tf
+
+from .base_model import BaseModel
+from .cnn_wrapper.aslfeat import ASLFeatNet
+
+sys.path.append('..')
+
+
+class FeatModel(BaseModel):
+    output_tensors = ["desc:0", "kpt:0", "score:0"]
+    default_config = {'max_dim': 1280}
+
+    def _init_model(self):
+        return
+
+    def _run(self, data):
+        assert len(data.shape) == 3
+        max_dim = max(data.shape[0], data.shape[1])
+        downsample_ratio = 1
+        if max_dim > self.config['max_dim']:
+            downsample_ratio = self.config['max_dim'] / float(max_dim)
+            data = cv2.resize(data, (0, 0), fx=downsample_ratio, fy=downsample_ratio)
+            data = data[..., np.newaxis]
+        feed_dict = {"input:0": np.expand_dims(data, 0)}
+        returns = self.sess.run(self.output_tensors, feed_dict=feed_dict)
+        desc = np.squeeze(returns[0], axis=0)
+        kpt = np.squeeze(returns[1], axis=0)
+        kpt /= downsample_ratio
+        score = np.squeeze(returns[2], axis=0)
+        return desc, kpt, score
+
+    def _construct_network(self):
+        ph_imgs = tf.placeholder(dtype=tf.float32, shape=(None, None, None, 1), name='input')
+        mean, variance = tf.nn.moments(
+            tf.cast(ph_imgs, tf.float32), axes=[1, 2], keep_dims=True)
+        norm_input = tf.nn.batch_normalization(ph_imgs, mean, variance, None, None, 1e-5)
+        config_dict = {'det_config': self.config['config']}
+        tower = ASLFeatNet({'data': norm_input}, is_training=False, resue=False, **config_dict)