FineNuImages - nuImages 2D detection in the unified detection format

Source: nuimages-v1.0-all-metadata.tgz + nuimages-v1.0-all-samples.tgz from the Motional AWS Open Data bucket (motional-nuscenes.s3.amazonaws.com), the same archives served by the nuscenes.org download page.

Converted by the finedet project into a unified, AutoTrain-compatible layout: image / width / height / objects{bbox, category} with COCO-format [x, y, w, h] boxes in absolute pixels. Boxes are clipped to the image and empty boxes dropped; category ids are densified per the category tables below.

Box format

objects.bbox follows the COCO convention: [x, y, w, h] in absolute pixels, origin at the image's top-left corner.

License

nuImages (Motional) is provided under CC BY-NC-SA 4.0 with additional terms — nuScenes Terms of Use: 'Unless specifically labeled otherwise, these Datasets are provided to You under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License ("CC BY-NC-SA 4.0"), with the additional terms included herein.' https://www.nuscenes.org/terms-of-use. This conversion is likewise distributed under CC BY-NC-SA 4.0. Data (c) Motional AD Inc. Motional does not sponsor, approve, or endorse this conversion.

Example images

Boxes are colored by category: near-transparent fill, opaque outline.

Conversion notes

Annotated key-frame camera images only (the +/-6-frame sweeps context is unannotated and not included). bbox converted from x1/y1/x2/y2 to COCO xywh. The native 23-class hierarchical foreground vocabulary is kept (no 10-class remap). Instance masks, object attributes, and the mask-only surface annotations (driveable surface, ego vehicle) are not carried over; a few stray object_ann rows labeled vehicle.ego (6 of ~694K) are dropped with them. v1.0-test has no public annotations and is excluded.

Splits

• train: 67279 images
• validation: 16445 images

Categories

id	name
0	animal
1	human.pedestrian.adult
2	human.pedestrian.child
3	human.pedestrian.construction_worker
4	human.pedestrian.personal_mobility
5	human.pedestrian.police_officer
6	human.pedestrian.stroller
7	human.pedestrian.wheelchair
8	movable_object.barrier
9	movable_object.debris
10	movable_object.pushable_pullable
11	movable_object.trafficcone
12	static_object.bicycle_rack
13	vehicle.bicycle
14	vehicle.bus.bendy
15	vehicle.bus.rigid
16	vehicle.car
17	vehicle.construction
18	vehicle.emergency.ambulance
19	vehicle.emergency.police
20	vehicle.motorcycle
21	vehicle.trailer
22	vehicle.truck

Training with transformers

The boxes are already in the absolute-pixel COCO [x, y, w, h] format that AutoImageProcessor expects, so fine-tuning a detector needs no bbox conversion:

import torch
from datasets import load_dataset
from transformers import (AutoImageProcessor, AutoModelForObjectDetection,
                          Trainer, TrainingArguments)

ds = load_dataset("finedet/nuimages")
obj_feat = ds["train"].features["objects"]
if hasattr(obj_feat, "feature"):
    obj_feat = obj_feat.feature
cat_feat = obj_feat["category"]
names = (cat_feat.feature if hasattr(cat_feat, "feature") else cat_feat).names

checkpoint = "facebook/detr-resnet-50"
processor = AutoImageProcessor.from_pretrained(checkpoint)
model = AutoModelForObjectDetection.from_pretrained(
    checkpoint,
    id2label=dict(enumerate(names)),
    label2id={n: i for i, n in enumerate(names)},
    ignore_mismatched_sizes=True,
)


def transform(batch):
    images = [img.convert("RGB") for img in batch["image"]]
    annotations = [
        {"image_id": i,
         "annotations": [
             {"bbox": box, "category_id": cat, "area": box[2] * box[3], "iscrowd": 0}
             for box, cat in zip(objs["bbox"], objs["category"])
         ]}
        for i, objs in enumerate(batch["objects"])
    ]
    return processor(images=images, annotations=annotations, return_tensors="pt")


def collate(batch):
    return {"pixel_values": torch.stack([x["pixel_values"] for x in batch]),
            "labels": [x["labels"] for x in batch]}


trainer = Trainer(
    model=model,
    args=TrainingArguments(output_dir="out", per_device_train_batch_size=4,
                           num_train_epochs=10, learning_rate=1e-5,
                           remove_unused_columns=False),
    train_dataset=ds["train"].with_transform(transform),
    data_collator=collate,
)
trainer.train()