act13/read_the_object

Read The Object

Read The Object is the official data release accompanying the paper "Read the Object: A Dataset and Task for Text-Grounded Object Understanding in Real-World Images" (Chowdhury, Park, Le, Chao, Hayat, Porikli, Mahajan — The Ohio State University, Qualcomm AI Research, Boston University, York University, Vector Institute for AI).

Objects in the real world are often distinguished not by visual appearance alone but by the text printed on them — a traffic sign's route number, a jersey number, a product label. The paper introduces text–object grounding: the task of associating a piece of scene text with the exact object instance it appears on, and shows that current object detectors, OCR systems, and vision-language models all struggle at it despite doing well on text recognition or object localization in isolation. The full benchmark combines images from OpenImages V5, MS COCO, and nuScenes (6,000 images, 25,208 referring-expression samples, 34,085 object instances, 137,246 text instances) and defines six complementary tasks: Referring Expression Comprehension, Text–Object Grounding, OCR, Text-Region Grounding, Visual Question Answering, and Multi-Text Captioning.

This repo currently contains only the nuScenes source category of that benchmark (the OpenImages V5 and COCO portions, plus the VQA and Captioning tasks, are not yet part of this HF release — more will be added incrementally). This is benchmark/evaluation data, not a training split — every config below ships only a test split.

The dataset is organized by source, then by task, so future source additions (e.g. COCO) slot in as sibling folders under the same layout:

read_the_object/
  <source>/
    images/                    clean source images (shared across tasks)
    referring_expression/      REC samples: referring_expression.json + images_bbox/
    captioning/                Captioning samples: captioning.json
    text_object_grounding/     TOG samples: text_object_grounding.json + images_bbox/
    text_region_grounding/     TRG samples: text_region_grounding.json + images_bbox/
    ocr/                       OCR samples: ocr.json + images_bbox/

Each images_bbox/ image is drawn per-sample (red = target object box, cyan = the specific text box(es) that sample's row refers to) — not every box in the source image, only the ones relevant to that row. Captioning has no images_bbox/, since its output is a whole-image caption rather than a box.

For datasets/Dataset Viewer users, each task is also published as its own Parquet config (embedded images, test split):

from datasets import load_dataset
ds = load_dataset("act13/read_the_object", "nuscenes_referring_expression", split="test")

config_name	task	rows
nuscenes_referring_expression	Referring Expression Comprehension	1,735
nuscenes_captioning	Multi-Text Captioning	259
nuscenes_text_object_grounding	Text–Object Grounding	2,107
nuscenes_text_region_grounding	Text–Region Grounding	2,140
nuscenes_ocr	OCR	259

nuscenes/referring_expression/referring_expression.json — Referring Expression Comprehension

Given a natural-language expression that includes both text and visual attributes, locate the described object and its text region. One row per referring-expression sample.

field	type	description
id	string	unique sample id, e.g. nuscenes_00000
source	string	"nuscenes"
image	string	relative path to the clean image
image_bbox	string	relative path to the same image with just this sample's box(es) drawn
category	string	object category, supercategory.subtype (e.g. vehicle.bus, signage.traffic_sign)
obj_box	[x1,y1,x2,y2]	target object's bounding box, pixel coords
texts	[string, ...]	the scene-text string(s) used to build the expression
combined_text_box	[x1,y1,x2,y2]	bounding box of the text region(s) used (always contained within obj_box)
ref_expression	string	the referring expression, e.g. "the bus with 'stop' written on it"
num_texts_used	int	how many text instances were combined into this sample
sample_type	string	single_text_single_object, multi_text_single_object, or disambiguated_same_category
attribute	string (optional)	present only for disambiguated_same_category samples — a distinguishing visual/spatial attribute added because another object of the same category in the same image shares the same text (e.g. "left blue" in "'23' written on left blue gate")

nuscenes/captioning/captioning.json — Multi-Text Captioning

Generate a caption that jointly describes the objects in the image and the visible text on them. One row per image, 259 total. Captions are 20-100 words, generated by grounding on the same verified object/text annotations as the other tasks, then manually spot-checked against the source images for accuracy.

field	type	description
id	string	unique sample id, e.g. nuscenes_caption_00000
source	string	"nuscenes"
image	string	relative path to the clean image
caption	string	a 20-100 word caption describing the scene and weaving in the visible scene-text

nuscenes/text_object_grounding/text_object_grounding.json — Text–Object Grounding

Given a text string, find all objects in the image that contain or are associated with that text. One row per (image, query_text) pair; if the same text is shared by multiple objects, all of them are listed (and all highlighted in image_bbox).

field	type	description
id	string	unique sample id, e.g. nuscenes_tog_00000
source	string	"nuscenes"
image	string	relative path to the clean image
image_bbox	string	relative path to the image with the query text box(es) (cyan) and every linked object box (red) drawn
query_text	string	the text string being grounded
categories	[string, ...]	category of each object this text is linked to
obj_boxes	[[x1,y1,x2,y2], ...]	box for each entry in categories, same order

nuscenes/text_region_grounding/text_region_grounding.json — Text–Region Grounding

Given a query text string, localize the corresponding text region(s) in the image. One row per (image, query_text) pair; if the same text appears multiple times in the image, every occurrence is listed (and all highlighted in image_bbox).

field	type	description
id	string	unique sample id, e.g. nuscenes_trg_00000
source	string	"nuscenes"
image	string	relative path to the clean image
image_bbox	string	relative path to the image with every occurrence of the query text boxed (cyan)
query_text	string	the text string being localized
boxes	[[x1,y1,x2,y2], ...]	every box where query_text appears in this image

nuscenes/ocr/ocr.json — OCR

Detect and recognize all visible text regions in the image. One row per image; image_bbox shows every verified text box in that image.

field	type	description
id	string	unique sample id, e.g. nuscenes_ocr_00000
source	string	"nuscenes"
image	string	relative path to the clean image
image_bbox	string	relative path to the image with all text boxes drawn
texts	[string, ...]	transcription of each text region in the image
boxes	[[x1,y1,x2,y2], ...]	box for each entry in texts, same order