Add nuScenes source: images, bbox images, and REC annotations

README.md

@@ -0,0 +1,125 @@
+---
+license: cc-by-nc-sa-4.0
+task_categories:
+- zero-shot-object-detection
+- image-to-text
+language:
+- en
+tags:
+- referring-expression-comprehension
+- visual-grounding
+- ocr
+- scene-text
+- autonomous-driving
+- nuscenes
+pretty_name: Read The Object
+size_categories:
+- 1K<n<10K
+---
+
+# Read The Object
+
+**Read The Object** is a Referring Expression Comprehension (REC) dataset that
+grounds *scene text* to the physical object it is written on, built from
+driving-scene imagery. Each sample pairs a natural-language referring
+expression (built from an object's category and the text detected on it)
+with the object's bounding box.
+
+The dataset is organized by source — the first category is **nuScenes**,
+with more (starting with COCO) planned. Each source lives in its own
+top-level folder with the same layout:
+
+```
+read_the_object/
+  <source>/
+    images/            clean source images
+    images_bbox/       same images with boxes drawn (red = object, cyan = text)
+    annotations.json   REC samples for this source
+```
+
+## `nuscenes/annotations.json`
+
+One row per referring-expression sample (**1,735** total, over **259**
+unique images from nuScenes v1.0-trainval `samples/` camera frames —
+CAM_FRONT, CAM_BACK_LEFT, CAM_BACK_RIGHT).
+
+| 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 boxes 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"`) |
+
+### Sample types
+
+- **`single_text_single_object`** — one object, one piece of text.
+- **`multi_text_single_object`** — one object with multiple pieces of text on
+  it; one sample is generated per count *k* = 1..5 using the *k* spatially
+  closest texts, so the same object contributes several samples of
+  increasing expression detail.
+- **`disambiguated_same_category`** — used only when two or more objects of
+  the **same category** in the same image share the exact same text string
+  (so category + text alone would be ambiguous). A human-annotated
+  distinguishing attribute (color/material/shape and/or left/right/top/center)
+  is folded into the expression: `'{text}' written on {attribute} {category}`.
+  Cross-category sharing (e.g. a gate and a building both labeled `"303"`)
+  is **not** treated as ambiguous, since the category name itself already
+  disambiguates those.
+
+## Construction pipeline
+
+1. **OCR**: EasyOCR and PaddleOCR run independently on every camera frame.
+   A text box is kept only if both models agree exactly on the recognized
+   string, both confidences exceed 0.7, and the IoU between their two boxes
+   exceeds 0.7.
+2. **Scene-aware temporal dedup**: nuScenes scenes are ~20s clips at 2Hz;
+   the same physical text is visible across many consecutive frames. Runs
+   of the same `(scene, camera, text)` are collapsed to one representative
+   frame per spatially-contiguous run.
+3. **3D object association**: each nuScenes 3D object annotation is
+   projected into the camera image (ego/global frame transforms + camera
+   intrinsics); a text box is linked to an object if the object's projected
+   silhouette contains it.
+4. **Manual review**: every image with 2+ text detections was manually
+   reviewed and corrected (text/object boxes fixed, text-to-object links
+   verified) via a custom annotation tool.
+5. **Categorization**: free-text object names were mapped to a fixed
+   category taxonomy (`vehicle.*`, `signage.*`, `structure.*`, `barrier.*`,
+   `misc.*`).
+6. **REC sample construction**: samples generated per the rules above, with
+   every `combined_text_box` guaranteed to be fully contained in its
+   `obj_box`.
+7. **Ambiguity resolution**: same-category ties (a shared text string across
+   2+ objects of the same category) were manually annotated with a
+   distinguishing attribute so a unique referring expression is always
+   possible.
+
+## Source data & license
+
+Images are camera frames from the [nuScenes](https://www.nuscenes.org/)
+dataset (© nuTonomy / Motional), used under the nuScenes non-commercial
+license. This derived annotation set (`annotations.json`, box annotations,
+categories, and referring expressions) follows the same non-commercial
+terms (`cc-by-nc-sa-4.0`). Redistribution of the underlying nuScenes images
+is subject to nuScenes' own terms — see https://www.nuscenes.org/terms-of-use.
+
+## Citation
+
+If you use this dataset, please cite nuScenes:
+
+```
+@inproceedings{nuscenes2019,
+  title={nuScenes: A multimodal dataset for autonomous driving},
+  author={Caesar, Holger and Bankiti, Varun and Lang, Alex H. and others},
+  booktitle={CVPR},
+  year={2020}
+}
+```