aggregate captions

LocalizedNarratives.py

@@ -60,33 +60,45 @@ _ANNOTATION_URLs = {
 }
 
 
-_FEATURES = datasets.Features(
-    {
-        "image": datasets.Image(),
-        "image_url": datasets.Value("string"),
-        "dataset_id": datasets.Value("string"),
-        "image_id": datasets.Value("string"),
-        "annotator_id": datasets.Value("int32"),
-        "caption": datasets.Value("string"),
-        "timed_caption": datasets.Sequence(
-            {
-                "utterance": datasets.Value("string"),
-                "start_time": datasets.Value("float32"),
-                "end_time": datasets.Value("float32"),
-            }
-        ),
-        "traces": datasets.Sequence(
-            datasets.Sequence(
+_FEATURES = {
+    "OpenImages": datasets.Features(
+        {
+            "image": datasets.Image(),
+            "image_url": datasets.Value("string"),
+            "dataset_id": datasets.Value("string"),
+            "image_id": datasets.Value("string"),
+            "annotator_id": datasets.Value("int32"),
+            "caption": datasets.Value("string"),
+            "timed_caption": datasets.Sequence(
                 {
-                    "x": datasets.Value("float32"),
-                    "y": datasets.Value("float32"),
-                    "t": datasets.Value("float32"),
+                    "utterance": datasets.Value("string"),
+                    "start_time": datasets.Value("float32"),
+                    "end_time": datasets.Value("float32"),
                 }
-            )
-        ),
-        "voice_recording": datasets.Value("string"),
-    }
-)
+            ),
+            "traces": datasets.Sequence(
+                datasets.Sequence(
+                    {
+                        "x": datasets.Value("float32"),
+                        "y": datasets.Value("float32"),
+                        "t": datasets.Value("float32"),
+                    }
+                )
+            ),
+            "voice_recording": datasets.Value("string"),
+        }
+    ),
+    "OpenImages_captions": datasets.Features(
+        {
+            "image": datasets.Image(),
+            "image_url": datasets.Value("string"),
+            "dataset_id": datasets.Value("string"),
+            "image_id": datasets.Value("string"),
+            "annotator_ids": [datasets.Value("int32")],
+            "captions": [datasets.Value("string")],
+        }
+    ),
+}
 
 
 class LocalizedNarrativesOpenImages(datasets.GeneratorBasedBuilder):
@@ -95,7 +107,16 @@ class LocalizedNarrativesOpenImages(datasets.GeneratorBasedBuilder):
     VERSION = datasets.Version("1.0.0")
 
     BUILDER_CONFIGS = [
-        datasets.BuilderConfig(name="OpenImages", version=VERSION, description="OpenImages subset of Localized Narratives"),
+        datasets.BuilderConfig(
+            name="OpenImages",
+            version=VERSION,
+            description="OpenImages subset of Localized Narratives"
+        ),
+        datasets.BuilderConfig(
+            name="OpenImages_captions",
+            version=VERSION,
+            description="OpenImages subset of Localized Narratives where captions are groupped per image (images can have multiple captions). For this subset, `timed_caption`, `traces` and `voice_recording` are not available."
+        ),
     ]
 
     DEFAULT_CONFIG_NAME = "OpenImages"
@@ -103,7 +124,7 @@ class LocalizedNarrativesOpenImages(datasets.GeneratorBasedBuilder):
     def _info(self):
         return datasets.DatasetInfo(
             description=_DESCRIPTION,
-            features=_FEATURES,
+            features=_FEATURES[self.config.name],
             homepage=_HOMEPAGE,
             license=_LICENSE,
             citation=_CITATION,
@@ -120,6 +141,12 @@ class LocalizedNarrativesOpenImages(datasets.GeneratorBasedBuilder):
         ]
 
     def _generate_examples(self, annotation_list: str, split: str):
+        if self.config.name == "OpenImages":
+            return self._generate_examples_original_format(annotation_list, split)
+        elif self.config.name == "OpenImages_captions":
+            return self._generate_examples_aggregated_captions(annotation_list, split)
+
+    def _generate_examples_original_format(self, annotation_list: str, split: str):
         counter = 0
         for annotation_file in annotation_list:
             with open(annotation_file, "r", encoding="utf-8") as fi:
@@ -138,3 +165,30 @@ class LocalizedNarrativesOpenImages(datasets.GeneratorBasedBuilder):
                         "voice_recording": annotation["voice_recording"],
                     }
                     counter += 1
+
+    def _generate_examples_aggregated_captions(self, annotation_list: str, split: str):
+        result = {}
+        for annotation_file in annotation_list:
+            with open(annotation_file, "r", encoding="utf-8") as fi:
+                for line in fi:
+                    annotation = json.loads(line)
+                    image_url = f"https://s3.amazonaws.com/open-images-dataset/{split}/{annotation['image_id']}.jpg"
+                    image_id = annotation["image_id"]
+                    if image_id in result:
+                        assert result[image_id]["dataset_id"] == annotation["dataset_id"]
+                        assert result[image_id]["image_id"] == annotation["image_id"]
+                        result[image_id]["annotator_ids"].append(annotation["annotator_id"])
+                        result[image_id]["captions"].append(annotation["caption"])
+                    else:
+                        result[image_id] = {
+                            "image": image_url,
+                            "image_url": image_url,
+                            "dataset_id": annotation["dataset_id"],
+                            "image_id": image_id,
+                            "annotator_ids": [annotation["annotator_id"]],
+                            "captions": [annotation["caption"]],
+                        }
+        counter = 0
+        for r in result.values():
+            yield counter, r
+            counter += 1

README.md

@@ -45,7 +45,9 @@ Since the voice and the mouse pointer are synchronized, we can localize every si
 This dense visual grounding takes the form of a mouse trace segment per word and is unique to our data.
 We annotated 849k images with Localized Narratives: the whole COCO, Flickr30k, and ADE20K datasets, and 671k images of Open Images, all of which we make publicly available.
 
-As of now, there is only the OpenImages subset, but feel free to contribute the other subset of Localized Narratives!
+As of now, there is only the `OpenImages` subset, but feel free to contribute the other subset of Localized Narratives!
+
+`OpenImages_captions` is similar to the `OpenImages` subset. The differences are that captions are groupped per image (images can have multiple captions). For this subset, `timed_caption`, `traces` and `voice_recording` are not available.
 
 ### Supported Tasks and Leaderboards