Merge pull request #8768 from kmindspark:newavarecords

PiperOrigin-RevId: 328778119

research/object_detection/dataset_tools/download_and_preprocess_ava.sh

+#!/bin/bash
+# This script downloads the videos for the AVA dataset. There are no arguments.
+# Copy this script into the desired parent directory of the ava_vids_raw/
+# directory created in this script to store the raw videos.
+
+mkdir ava_vids_raw
+cd ava_vids_raw
+
+curl -O s3.amazonaws.com/ava-dataset/annotations/ava_file_names_trainval_v2.1.txt
+
+echo "Downloading all videos."
+
+cat "ava_file_names_trainval_v2.1.txt" | while read line
+do
+  curl -O s3.amazonaws.com/ava-dataset/trainval/$line
+  echo "Downloaded " $line
+done
+
+rm "ava_file_names_trainval_v2.1.txt"
+cd ..
+
+# Trimming causes issues with frame seeking in the python script, so it is best left out.
+# If included, need to modify the python script to subtract 900 seconds wheen seeking.
+
+# echo "Trimming all videos."
+
+# mkdir ava_vids_trimmed
+# for filename in ava_vids_raw/*; do
+#   ffmpeg -ss 900 -to 1800 -i $filename -c copy ava_vids_trimmed/${filename##*/}
+# done

research/object_detection/dataset_tools/seq_example_util.py

-# Lint as: python2, python3
 # Copyright 2020 The TensorFlow Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -123,6 +122,15 @@ def sequence_bytes_feature(ndarray):
   return feature_list
 
 
+def sequence_strings_feature(strings):
+  new_str_arr = []
+  for single_str in strings:
+    new_str_arr.append(tf.train.Feature(
+        bytes_list=tf.train.BytesList(
+            value=[single_str.encode('utf8')])))
+  return tf.train.FeatureList(feature=new_str_arr)
+
+
 def boxes_to_box_components(bboxes):
   """Converts a list of numpy arrays (boxes) to box components.
 
@@ -137,8 +145,11 @@ def boxes_to_box_components(bboxes):
   ymax_list = []
   xmax_list = []
   for bbox in bboxes:
-    bbox = np.array(bbox).astype(np.float32)
-    ymin, xmin, ymax, xmax = np.split(bbox, 4, axis=1)
+    if bbox != []:  # pylint: disable=g-explicit-bool-comparison
+      bbox = np.array(bbox).astype(np.float32)
+      ymin, xmin, ymax, xmax = np.split(bbox, 4, axis=1)
+    else:
+      ymin, xmin, ymax, xmax = [], [], [], []
     ymin_list.append(np.reshape(ymin, [-1]))
     xmin_list.append(np.reshape(xmin, [-1]))
     ymax_list.append(np.reshape(ymax, [-1]))
@@ -159,7 +170,8 @@ def make_sequence_example(dataset_name,
                           label_strings=None,
                           detection_bboxes=None,
                           detection_classes=None,
-                          detection_scores=None):
+                          detection_scores=None,
+                          use_strs_for_source_id=False):
   """Constructs tf.SequenceExamples.
 
   Args:
@@ -189,6 +201,8 @@ def make_sequence_example(dataset_name,
     detection_scores: (Optional) A list (with num_frames_elements) of
       [num_boxes_i] numpy float32 arrays holding predicted object scores for
       each frame.
+    use_strs_for_source_id: (Optional) Whether to write the source IDs as
+      strings rather than byte lists of characters.
 
   Returns:
     A tf.train.SequenceExample.
@@ -221,7 +235,11 @@ def make_sequence_example(dataset_name,
   if image_format is not None:
     context_dict['image/format'] = context_bytes_feature([image_format])
   if image_source_ids is not None:
-    feature_list['image/source_id'] = sequence_bytes_feature(image_source_ids)
+    if use_strs_for_source_id:
+      feature_list['image/source_id'] = sequence_strings_feature(
+          image_source_ids)
+    else:
+      feature_list['image/source_id'] = sequence_bytes_feature(image_source_ids)
   if bboxes is not None:
     bbox_ymin, bbox_xmin, bbox_ymax, bbox_xmax = boxes_to_box_components(bboxes)
     feature_list['region/bbox/xmin'] = sequence_float_feature(bbox_xmin)

research/object_detection/dataset_tools/seq_example_util_test.py

@@ -104,12 +104,12 @@ class SeqExampleUtilTest(tf.test.TestCase):
         source_ids)
 
   def test_make_labeled_example(self):
-    num_frames = 2
+    num_frames = 3
     image_height = 100
     image_width = 200
     dataset_name = b'unlabeled_dataset'
     video_id = b'video_000'
-    labels = [b'dog', b'cat']
+    labels = [b'dog', b'cat', b'wolf']
     images = tf.cast(tf.random.uniform(
         [num_frames, image_height, image_width, 3],
         maxval=256,
@@ -117,15 +117,17 @@ class SeqExampleUtilTest(tf.test.TestCase):
     images_list = tf.unstack(images, axis=0)
     encoded_images_list = [tf.io.encode_jpeg(image) for image in images_list]
     encoded_images = self.materialize_tensors(encoded_images_list)
-    timestamps = [100000, 110000]
-    is_annotated = [1, 0]
+    timestamps = [100000, 110000, 120000]
+    is_annotated = [1, 0, 1]
     bboxes = [
         np.array([[0., 0., 0., 0.],
                   [0., 0., 1., 1.]], dtype=np.float32),
-        np.zeros([0, 4], dtype=np.float32)
+        np.zeros([0, 4], dtype=np.float32),
+        np.array([], dtype=np.float32)
     ]
     label_strings = [
         np.array(labels),
+        np.array([]),
         np.array([])
     ]