megadetector 10.0.15__py3-none-any.whl

megadetector/data_management/lila/generate_lila_per_image_labels.py

@@ -0,0 +1,777 @@
+"""
+
+generate_lila_per_image_labels.py
+
+Generate a .csv file with one row per annotation, containing full URLs to every
+camera trap image on LILA, with taxonomically expanded labels.
+
+Typically there will be one row per image, though images with multiple annotations
+will have multiple rows.
+
+Some images may not physically exist, particularly images that are labeled as "human".
+This script does not validate image URLs.
+
+Does not include bounding box annotations.
+
+"""
+
+#%% Constants and imports
+
+import os
+import json
+import pandas as pd
+import numpy as np
+import dateparser # type: ignore
+import csv
+
+from collections import defaultdict
+from tqdm import tqdm
+
+from megadetector.data_management.lila.lila_common import \
+    read_lila_metadata, \
+    read_metadata_file_for_dataset, \
+    read_lila_taxonomy_mapping
+
+from megadetector.utils import write_html_image_list
+from megadetector.utils.path_utils import zip_file
+from megadetector.utils.path_utils import open_file
+from megadetector.utils.url_utils import parallel_download_urls
+
+# We'll write images, metadata downloads, and temporary files here
+lila_local_base = os.path.expanduser('~/lila')
+preview_folder = os.path.join(lila_local_base,'csv_preview')
+
+os.makedirs(lila_local_base,exist_ok=True)
+
+metadata_dir = os.path.join(lila_local_base,'metadata')
+os.makedirs(metadata_dir,exist_ok=True)
+
+output_file = os.path.join(lila_local_base,'lila_image_urls_and_labels.csv')
+
+# Some datasets don't have "sequence_level_annotation" fields populated, but we know their
+# annotation level
+ds_name_to_annotation_level = {}
+ds_name_to_annotation_level['Caltech Camera Traps'] = 'image'
+ds_name_to_annotation_level['ENA24'] = 'image'
+ds_name_to_annotation_level['Island Conservation Camera Traps'] = 'image'
+ds_name_to_annotation_level['Channel IslandsCamera Traps'] = 'image'
+ds_name_to_annotation_level['WCS Camera Traps'] = 'sequence'
+ds_name_to_annotation_level['Wellington Camera Traps'] = 'sequence'
+ds_name_to_annotation_level['NACTI'] = 'unknown'
+ds_name_to_annotation_level['Seattle(ish) Camera Traps'] = 'image'
+
+known_unmapped_labels = set(['WCS Camera Traps:#ref!'])
+
+debug_max_images_per_dataset = -1
+if debug_max_images_per_dataset > 0:
+    print('Running in debug mode')
+    output_file = output_file.replace('.csv','_debug.csv')
+
+taxonomy_levels_to_include = \
+    ['kingdom','phylum','subphylum','superclass','class','subclass','infraclass','superorder','order',
+     'suborder','infraorder','superfamily','family','subfamily','tribe','genus','subgenus',
+     'species','subspecies','variety']
+
+def _clearnan(v):
+    if isinstance(v,float):
+        assert np.isnan(v)
+        v = ''
+    assert isinstance(v,str)
+    return v
+
+
+#%% Download and parse the metadata file
+
+metadata_table = read_lila_metadata(metadata_dir)
+
+# To select an individual data set for debugging
+if False:
+    k = 'Idaho Camera Traps'
+    metadata_table = {k:metadata_table[k]}
+
+
+#%% Download and extract metadata for each dataset
+
+for ds_name in metadata_table.keys():
+    metadata_table[ds_name]['metadata_filename'] = read_metadata_file_for_dataset(ds_name=ds_name,
+                                                                         metadata_dir=metadata_dir,
+                                                                         metadata_table=metadata_table)
+
+#%% Load taxonomy data
+
+taxonomy_df = read_lila_taxonomy_mapping(metadata_dir, force_download=True)
+
+
+#%% Build a dictionary that maps each [dataset,query] pair to the full taxonomic label set
+
+ds_label_to_taxonomy = {}
+
+# i_row = 0; row = taxonomy_df.iloc[i_row]
+for i_row,row in taxonomy_df.iterrows():
+
+    ds_label = row['dataset_name'] + ':' + row['query']
+    assert ds_label.strip() == ds_label
+    assert ds_label not in ds_label_to_taxonomy
+    ds_label_to_taxonomy[ds_label] = row.to_dict()
+
+
+#%% Process annotations for each dataset
+
+# Takes a few hours
+
+# The order of these headers needs to match the order in which fields are added later in this cell;
+# don't mess with this order.
+header = ['dataset_name','url_gcp','url_aws','url_azure',
+          'image_id','sequence_id','location_id','frame_num',
+          'original_label','scientific_name','common_name','datetime','annotation_level']
+
+header.extend(taxonomy_levels_to_include)
+
+missing_annotations = set()
+
+with open(output_file,'w',encoding='utf-8',newline='') as f:
+
+    csv_writer = csv.writer(f)
+    csv_writer.writerow(header)
+
+    # ds_name = list(metadata_table.keys())[0]
+    for ds_name in metadata_table.keys():
+
+        if 'bbox' in ds_name:
+            print('Skipping bbox dataset {}'.format(ds_name))
+            continue
+
+        print('Processing dataset {}'.format(ds_name))
+
+        json_filename = metadata_table[ds_name]['metadata_filename']
+        with open(json_filename, 'r') as f:
+            data = json.load(f)
+
+        categories = data['categories']
+        for c in categories:
+            category_id_to_name = {c['id']:c['name'] for c in categories}
+
+        annotations = data['annotations']
+        images = data['images']
+
+        image_id_to_annotations = defaultdict(list)
+
+        # Go through annotations, marking each image with the categories that are present
+        #
+        # ann = annotations[0]
+        for ann in annotations:
+            image_id_to_annotations[ann['image_id']].append(ann)
+
+        unannotated_images = []
+
+        found_date = False
+        found_location = False
+        found_annotation_level = False
+
+        if ds_name in ds_name_to_annotation_level:
+            expected_annotation_level = ds_name_to_annotation_level[ds_name]
+        else:
+            expected_annotation_level = None
+
+        # im = images[10]
+        for i_image,im in tqdm(enumerate(images),total=len(images)):
+
+            if (debug_max_images_per_dataset is not None) and (debug_max_images_per_dataset > 0) \
+                and (i_image >= debug_max_images_per_dataset):
+                break
+
+            file_name = im['file_name'].replace('\\','/')
+            base_url_gcp = metadata_table[ds_name]['image_base_url_gcp']
+            base_url_aws = metadata_table[ds_name]['image_base_url_aws']
+            base_url_azure = metadata_table[ds_name]['image_base_url_azure']
+            assert not base_url_gcp.endswith('/')
+            assert not base_url_aws.endswith('/')
+            assert not base_url_azure.endswith('/')
+
+            url_gcp = base_url_gcp + '/' + file_name
+            url_aws = base_url_aws + '/' + file_name
+            url_azure = base_url_azure + '/' + file_name
+
+            for k in im.keys():
+                if ('date' in k or 'time' in k) and (k not in ['datetime','date_captured']):
+                    raise ValueError('Unrecognized datetime field')
+
+            # This field name was only used for Caltech Camera Traps
+            if 'date_captured' in im:
+                assert ds_name == 'Caltech Camera Traps'
+                im['datetime'] = im['date_captured']
+
+            def _has_valid_datetime(im):
+                if 'datetime' not in im:
+                    return False
+                v = im['datetime']
+                if v is None:
+                    return False
+                if isinstance(v,str):
+                    return len(v) > 0
+                else:
+                    assert isinstance(v,float) and np.isnan(v)
+                    return False
+
+            dt_string = ''
+            if (_has_valid_datetime(im)):
+
+                dt = dateparser.parse(im['datetime'])
+
+                if dt is None or dt.year < 1990 or dt.year > 2025:
+
+                    # raise ValueError('Suspicious date parsing result')
+
+                    # Special case we don't want to print a warning about... this is
+                    # in invalid date that very likely originates on the camera, not at
+                    # some intermediate processing step.
+                    #
+                    # print('Suspicious date for image {}: {} ({})'.format(
+                    #    im['id'], im['datetime'], ds_name))
+                    pass
+
+                else:
+
+                    found_date = True
+                    dt_string = dt.strftime("%m-%d-%Y %H:%M:%S")
+
+            # Location, sequence, and image IDs are only guaranteed to be unique within
+            # a dataset, so for the output .csv file, include both
+            if 'location' in im:
+                found_location = True
+                location_id = ds_name + ' : ' + str(im['location'])
+            else:
+                location_id = ds_name
+
+            image_id = ds_name + ' : ' + str(im['id'])
+
+            if 'seq_id' in im:
+                sequence_id = ds_name + ' : ' + str(im['seq_id'])
+            else:
+                sequence_id = ds_name + ' : ' + 'unknown'
+
+            if 'frame_num' in im:
+                frame_num = im['frame_num']
+            else:
+                frame_num = -1
+
+            annotations_this_image = image_id_to_annotations[im['id']]
+
+            categories_this_image = set()
+
+            annotation_level = 'unknown'
+
+            for ann in annotations_this_image:
+                assert ann['image_id'] == im['id']
+                categories_this_image.add(category_id_to_name[ann['category_id']])
+                if 'sequence_level_annotation' in ann:
+                    found_annotation_level = True
+                    if ann['sequence_level_annotation']:
+                        annotation_level = 'sequence'
+                    else:
+                        annotation_level = 'image'
+                    if expected_annotation_level is not None:
+                        assert expected_annotation_level == annotation_level,\
+                            'Unexpected annotation level'
+                elif expected_annotation_level is not None:
+                    annotation_level = expected_annotation_level
+
+            if len(categories_this_image) == 0:
+                unannotated_images.append(im)
+                continue
+
+            # category_name = list(categories_this_image)[0]
+            for category_name in categories_this_image:
+
+                ds_label = ds_name + ':' + category_name.lower()
+
+                if ds_label not in ds_label_to_taxonomy:
+
+                    assert ds_label in known_unmapped_labels
+
+                    # Only print a warning the first time we see an unmapped label
+                    if ds_label not in missing_annotations:
+                        print('Warning: {} not in taxonomy file'.format(ds_label))
+                    missing_annotations.add(ds_label)
+                    continue
+
+                taxonomy_labels = ds_label_to_taxonomy[ds_label]
+
+                """
+                header =
+                    ['dataset_name','url','image_id','sequence_id','location_id',
+                     'frame_num','original_label','scientific_name','common_name',
+                     'datetime','annotation_level']
+                """
+
+                row = []
+                row.append(ds_name)
+                row.append(url_gcp)
+                row.append(url_aws)
+                row.append(url_azure)
+                row.append(image_id)
+                row.append(sequence_id)
+                row.append(location_id)
+                row.append(frame_num)
+                row.append(taxonomy_labels['query'])
+                row.append(_clearnan(taxonomy_labels['scientific_name']))
+                row.append(_clearnan(taxonomy_labels['common_name']))
+                row.append(dt_string)
+                row.append(annotation_level)
+
+                for s in taxonomy_levels_to_include:
+                    row.append(_clearnan(taxonomy_labels[s]))
+
+                assert len(row) == len(header)
+
+                csv_writer.writerow(row)
+
+            # ...for each category that was applied at least once to this image
+
+        # ...for each image in this dataset
+
+        if not found_date:
+            pass
+            # print('Warning: no date information available for this dataset')
+
+        if not found_location:
+            pass
+            # print('Warning: no location information available for this dataset')
+
+        if not found_annotation_level and (ds_name not in ds_name_to_annotation_level):
+            print('Warning: no annotation level information available for this dataset')
+
+        if len(unannotated_images) > 0:
+            print('Warning: {} of {} images are un-annotated\n'.\
+                  format(len(unannotated_images),len(images)))
+
+    # ...for each dataset
+
+# ...with open()
+
+print('\nProcessed {} datasets'.format(len(metadata_table)))
+
+
+#%% Read the .csv back
+
+df = pd.read_csv(output_file, low_memory=False)
+print('Read {} rows from {}'.format(len(df),output_file))
+
+
+#%% Do some post-hoc integrity checking
+
+# Takes ~5 minutes with apply(), or ~10 minutes without apply()
+#
+# Using apply() is faster, but more annoying to debug.
+use_pandas_apply_for_integrity_checking = True
+
+tqdm.pandas()
+
+def _isint(v):
+    return isinstance(v,int) or isinstance(v,np.int64)
+
+valid_annotation_levels = set(['sequence','image','unknown'])
+
+# Collect a list of locations within each dataset; we'll use this
+# in the next cell to look for datasets that only have a single location
+dataset_name_to_locations = defaultdict(set)
+
+def _check_row(row):
+
+    assert row['dataset_name'] in metadata_table.keys()
+    for url_column in ['url_gcp','url_aws','url_azure']:
+        assert row[url_column].startswith('https://') or row[url_column].startswith('http://')
+    assert ' : ' in row['image_id']
+    assert 'seq' not in row['location_id'].lower()
+    assert row['annotation_level'] in valid_annotation_levels
+
+    # frame_num should either be NaN or an integer
+    if isinstance(row['frame_num'],float):
+        assert np.isnan(row['frame_num'])
+    else:
+        # -1 is sometimes used for sequences of unknown length
+        assert _isint(row['frame_num']) and row['frame_num'] >= -1
+
+    ds_name = row['dataset_name']
+    dataset_name_to_locations[ds_name].add(row['location_id'])
+
+if use_pandas_apply_for_integrity_checking:
+
+    df.progress_apply(_check_row, axis=1)
+
+else:
+
+    # i_row = 0; row = df.iloc[i_row]
+    for i_row,row in tqdm(df.iterrows(),total=len(df)):
+        _check_row(row)
+
+
+#%% Check for datasets that have only one location string (typically "unknown")
+
+# Expected: ENA24, Missouri Camera Traps, Desert Lion Conservation Camera Traps
+
+for ds_name in dataset_name_to_locations.keys():
+    if len(dataset_name_to_locations[ds_name]) == 1:
+        print('No location information for {}'.format(ds_name))
+
+
+#%% Preview constants
+
+n_empty_images_per_dataset = 3
+n_non_empty_images_per_dataset = 10
+
+os.makedirs(preview_folder,exist_ok=True)
+
+
+#%% Choose images to download
+
+# Takes ~60 seconds
+
+np.random.seed(0)
+images_to_download = []
+
+# ds_name = list(metadata_table.keys())[2]
+for ds_name in metadata_table.keys():
+
+    if 'bbox' in ds_name:
+        continue
+
+    # Find all rows for this dataset
+    ds_rows = df.loc[df['dataset_name'] == ds_name]
+
+    print('{} rows available for {}'.format(len(ds_rows),ds_name))
+    assert len(ds_rows) > 0
+
+    empty_rows = ds_rows[ds_rows['scientific_name'].isnull()]
+    non_empty_rows = ds_rows[~ds_rows['scientific_name'].isnull()]
+
+    if len(empty_rows) == 0:
+        print('No empty images available for {}'.format(ds_name))
+    elif len(empty_rows) > n_empty_images_per_dataset:
+        empty_rows = empty_rows.sample(n=n_empty_images_per_dataset)
+    images_to_download.extend(empty_rows.to_dict('records'))
+
+    # All LILA datasets have non-empty images
+    if len(non_empty_rows) == 0:
+        raise ValueError('No non-empty images available for {}'.format(ds_name))
+    elif len(non_empty_rows) > n_non_empty_images_per_dataset:
+        non_empty_rows = non_empty_rows.sample(n=n_non_empty_images_per_dataset)
+    images_to_download.extend(non_empty_rows.to_dict('records'))
+
+ # ...for each dataset
+
+print('Selected {} total images'.format(len(images_to_download)))
+
+
+#%% Download images (prep)
+
+# Expect a few errors for images with human or vehicle labels (or things like "ignore" that *could* be humans)
+
+preferred_cloud = 'gcp'
+
+url_to_target_file = {}
+
+# i_image = 10; image = images_to_download[i_image]
+for i_image,image in tqdm(enumerate(images_to_download),total=len(images_to_download)):
+
+    url = image['url_' + preferred_cloud]
+    ext = os.path.splitext(url)[1]
+    fn_relative = 'image_{}'.format(str(i_image).zfill(4)) + ext
+    fn_abs = os.path.join(preview_folder,fn_relative)
+    image['relative_file'] = fn_relative
+    image['url'] = url
+    url_to_target_file[url] = fn_abs
+
+
+#%% Download images (execution)
+
+download_results = parallel_download_urls(url_to_target_file,verbose=False,overwrite=True,
+                                          n_workers=20,pool_type='thread')
+
+# 10-20 errors is normal; they should all be images that are labeled as "human"
+errors = []
+
+for r in download_results:
+    if r['status'] != 'success':
+        errors.append(r)
+
+assert len(download_results) == len(url_to_target_file)
+print('Errors on {} of {} downloads:\n'.format(len(errors),len(download_results)))
+
+for err in errors:
+    print(err['url'])
+
+
+#%% Write preview HTML
+
+html_filename = os.path.join(preview_folder,'index.html')
+
+html_images = []
+
+# im = images_to_download[0]
+for im in images_to_download:
+
+    if im['relative_file'] is None:
+        continue
+
+    output_im = {}
+    output_im['filename'] = im['relative_file']
+    output_im['linkTarget'] = im['url']
+    output_im['title'] = '<b>{}: {}</b><br/><br/>'.format(im['dataset_name'],im['original_label']) + str(im)
+    output_im['imageStyle'] = 'width:600px;'
+    output_im['textStyle'] = 'font-weight:normal;font-size:100%;'
+    html_images.append(output_im)
+
+write_html_image_list.write_html_image_list(html_filename,html_images)
+
+open_file(html_filename)
+
+
+#%% Zip output file
+
+zipped_output_file = zip_file(output_file,verbose=True,overwrite=True)
+
+print('Zipped {} to {}'.format(output_file,zipped_output_file))
+
+
+#%% Experimental: convert to .json
+
+"""
+The .csv file "output_file" (already loaded into the variable "df" at this point) has the following columns:
+
+dataset_name,url_gcp,url_aws,url_azure,image_id,sequence_id,location_id,frame_num,original_label,scientific_name,common_name,datetime,annotation_level,kingdom,phylum,subphylum,superclass,class,subclass,infraclass,superorder,order,suborder,infraorder,superfamily,family,subfamily,tribe,genus,subgenus,species,subspecies,variety
+
+Each row in the .csv represents an image.  The URL columns represent the location of that
+image on three different clouds; for a given image, the value of those columns differs only
+in the prefix.  The columns starting with "kingdom" represent a taxonomic wildlife identifier.  Not
+all rows have values in all of these columns; some rows represent non-wildlife images where all of these
+columns are blank.
+
+This cell converts this to a .json dictionary, with the following top-level keys:
+
+## datasets (dict)
+
+A dict mapping integer IDs to strings.
+
+Each unique value in the "dataset_name" column should become an element in this dict with a unique ID.
+
+## sequences (dict)
+
+A dict mapping integer IDs to strings.
+
+Each unique value in the "sequence_id" column should become an element in this dict with a unique ID.
+
+## locations (dict)
+
+A dict mapping integer IDs to strings.
+
+Each unique value in the "location_id" column should become an element in this dict with a unique ID.
+
+## base_urls (dict)
+
+This key should point to the following dict:
+
+{
+"gcp": "https://storage.googleapis.com/public-datasets-lila/",
+"aws": "http://us-west-2.opendata.source.coop.s3.amazonaws.com/agentmorris/lila-wildlife/",
+"azure": "https://lilawildlife.blob.core.windows.net/lila-wildlife/",
+}
+
+All values in the url_gcp, url_aws, and url_azure columns start with these values, respectively.
+
+## taxa (dict)
+
+A dict mapping integer IDs to dicts, where each dict has the fields:
+
+kingdom,phylum,subphylum,superclass,class,subclass,infraclass,superorder,order,suborder,infraorder,superfamily,family,subfamily,tribe,genus,subgenus,species,subspecies,variety
+
+The value of each of these fields in each row is either a string or None.
+
+## images (list)
+
+A list of images, where each image is a dict with the following fields:
+
+### dataset (int)
+
+The integer ID corresponding to the dataset_name column for this image
+
+### path (str)
+
+The suffix for this image's URL, which should be the same across the three URL columns.
+
+### seq (int)
+
+The integer ID corresponding to the sequence_id column for this image
+
+### loc (int)
+
+The integer ID corresponding to the location_id column for this image
+
+### frame_num
+
+The value of the frame_num column for this image, unless the original value was -1,
+in which case this is omitted.
+
+### original_label
+
+The value of the original_label column for this image
+
+### common_name
+
+The value of the common_name column for this image, if not empty
+
+### datetime
+
+The value of the datetime column for this image
+
+### ann_level
+
+The value of the annotation_level column for this image
+
+### taxon
+
+The integer ID corresponding to the taxonomic identifier columns for this image
+
+--
+
+The original .csv file is large (~15GB); this may impact the implementation of the .json conversion.  Speed of
+conversion is not a priority.
+
+"""
+
+print('Converting to JSON...')
+
+output_json_file = output_file.replace('.csv', '.json')
+
+json_data = {}
+
+# Create mappings for datasets, sequences, and locations
+dataset_to_id = {}
+sequence_to_id = {}
+location_to_id = {}
+taxa_to_id = {}
+
+next_dataset_id = 0
+next_sequence_id = 0
+next_location_id = 0
+next_taxa_id = 0
+
+json_data['datasets'] = {}
+json_data['sequences'] = {}
+json_data['locations'] = {}
+json_data['taxa'] = {}
+
+json_data['base_urls'] = {
+    "gcp": "https://storage.googleapis.com/public-datasets-lila/",
+    "aws": "http://us-west-2.opendata.source.coop.s3.amazonaws.com/agentmorris/lila-wildlife/",
+    "azure": "https://lilawildlife.blob.core.windows.net/lila-wildlife/",
+}
+
+json_data['images'] = []
+
+debug_max_json_conversion_rows = None
+
+print('Counting rows in .csv file...')
+
+# Get total number of lines for progress bar (optional, but helpful for large files)
+def _count_lines(filename):
+    with open(filename, 'r', encoding='utf-8') as f:
+        return sum(1 for line in f) - 1
+
+total_rows = _count_lines(output_file)
+print('Total rows to process: {}'.format(total_rows))
+
+# Read CSV file line by line
+with open(output_file, 'r', encoding='utf-8') as csvfile:
+
+    reader = csv.DictReader(csvfile)
+
+    # Process each row
+    for i_row, row in enumerate(tqdm(reader, total=total_rows, desc="Processing rows")):
+
+        if (debug_max_json_conversion_rows is not None) and (i_row >= debug_max_json_conversion_rows):
+            break
+
+        # Datasets
+        dataset_name = row['dataset_name']
+        if dataset_name not in dataset_to_id:
+            dataset_to_id[dataset_name] = next_dataset_id
+            json_data['datasets'][str(next_dataset_id)] = dataset_name
+            next_dataset_id += 1
+        dataset_id = dataset_to_id[dataset_name]
+
+        # Sequences
+        sequence_id_str = row['sequence_id']
+        assert sequence_id_str.startswith(dataset_name + ' : ')
+        if sequence_id_str not in sequence_to_id:
+            sequence_to_id[sequence_id_str] = next_sequence_id
+            json_data['sequences'][str(next_sequence_id)] = sequence_id_str
+            next_sequence_id += 1
+        sequence_id = sequence_to_id[sequence_id_str]
+
+        # Locations
+        location_id_str = row['location_id']
+        assert location_id_str.startswith(dataset_name) # + ' : ')
+        if location_id_str not in location_to_id:
+            location_to_id[location_id_str] = next_location_id
+            json_data['locations'][str(next_location_id)] = location_id_str
+            next_location_id += 1
+        location_id = location_to_id[location_id_str]
+
+        # Taxa
+        taxa_data = {level: _clearnan(row[level]) for level in taxonomy_levels_to_include}
+        taxa_tuple = tuple(taxa_data.items())  # use tuple for hashable key
+        if taxa_tuple not in taxa_to_id:
+            taxa_to_id[taxa_tuple] = next_taxa_id
+            json_data['taxa'][str(next_taxa_id)] = taxa_data
+            next_taxa_id += 1
+        taxa_id = taxa_to_id[taxa_tuple]
+
+        # Image path
+        url_gcp = row['url_gcp']
+        assert url_gcp.startswith(json_data['base_urls']['gcp'])
+        path = url_gcp.replace(json_data['base_urls']['gcp'], '')
+
+        common_name = _clearnan(row['common_name'])
+
+        # Convert to float first in case this appears in the .csv file as, e.g. "3.0"
+        frame_num = int(float(row['frame_num']))
+
+        # Image data
+        image_entry = {
+            'dataset': dataset_id,
+            'path': path,
+            'seq': sequence_id,
+            'loc': location_id,
+            'ann_level': row['annotation_level'],
+            'original_label': row['original_label'],
+            'datetime': row['datetime'],
+            'taxon': taxa_id
+        }
+
+        if frame_num >= 0:
+           image_entry['frame_num'] = frame_num
+
+        if len(common_name) > 0:
+            image_entry['common_name'] = common_name
+
+        json_data['images'].append(image_entry)
+
+    # ...for each line
+
+# ...with open(...)
+
+# Save the JSON data
+print('Saving JSON file...')
+with open(output_json_file, 'w', encoding='utf-8') as f:
+    json.dump(json_data, f, indent=1)
+
+print(f'Converted to JSON and saved to {output_json_file}')
+print(f'JSON file size: {os.path.getsize(output_json_file)/(1024*1024*1024):.2f} GB')
+
+# Print summary statistics
+print(f'Total datasets: {len(json_data["datasets"])}')
+print(f'Total sequences: {len(json_data["sequences"])}')
+print(f'Total locations: {len(json_data["locations"])}')
+print(f'Total taxa: {len(json_data["taxa"])}')
+print(f'Total images: {len(json_data["images"])}')