How to explore and visualize ML data for object detection in images

In recent years, people have gained a deeper understanding of the importance of in-depth understanding of machine learning data (ML-data). However, since detecting large data sets usually requires a lot of human and material investment, its widespread application in the field of computer vision still requires further development.

Usually, in object detection (Object Detection, which is a subset of computer vision), objects in the image are positioned by defining bounding boxes, which can not only identify objects, but also Understand an object's context, size, and relationship to other elements in the scene. At the same time, a comprehensive understanding of the distribution of classes, the diversity of object sizes, and the common environments in which classes appear will also help to discover error patterns in the training model during evaluation and debugging, so that additional training data can be selected more targeted .

In practice, I often take the following approach:

• Use the enhanced functions of pre-trained models or basic models, Add structure to your data. For example: creating various image embeddings and employing dimensionality reduction techniques such as t-SNE or UMAP. These can generate similarity maps to facilitate data browsing. In addition, using pre-trained models for detection can also facilitate context extraction.
• Use visualization tools that can integrate such structures with statistical and review capabilities of the raw data.

Below, I will introduce how to use Renomics Spotlight to create interactive object detection visualizations. As an example, I will try to:

• Build a visualization for a person detector in an image.
• Visualizations include similarity maps, filters, and statistics for easy exploration of the data.
• View every image through ground truth and Ultralytics YOLOv8 detection details.

Goal visualization on Renomics Spotlight. Source: Created by the author

Download the character images in the COCO dataset

First, install the required software packages through the following command:

!pip install fiftyone ultralytics renumics-spotlight

Using

FiftyOne’s resumable download function, you can download various images from the COCO dataset. With simple parameter settings, we can download 1,000 images containing one or more people. The specific code is as follows:

importpandasaspdimportnumpyasnpimportfiftyone.zooasfoz# 从 COCO 数据集中下载 1000 张带人的图像dataset = foz.load_zoo_dataset( "coco-2017"、split="validation"、label_types=[ "detections"、],classes=["person"]、 max_samples=1000、dataset_name="coco-2017-person-1k-validations"、)

Next, you can use the following code:

def xywh_too_xyxyn(bbox)： "" convert from xywh to xyxyn format """ return[bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]].行 = []fori, samplein enumerate(dataset)：labels = [detection.labelfordetectioninsample.ground_truth.detections] bboxs = [...bboxs = [xywh_too_xyxyn(detection.bounding_box) fordetectioninsample.ground_truth.detections]bboxs_persons = [bboxforbbox, labelin zip(bboxs, labels)iflabel =="person"] 行。row.append([sample.filepath, labels, bboxs, bboxs_persons])df = pd.DataFrame(row, columns=["filepath","categories", "bboxs", "bboxs_persons"])df["major_category"] = df["categories"].apply( lambdax:max(set(x) -set(["person"]), key=x.count) if len(set(x)) >1 else "only person"。)

Prepare the data is a Pandas DataFrame with columns including: file path, bounding box category, bounding box, the person contained in the bounding box, and the main category (despite the person) to specify the context of the person in the image:

You can then visualize it via Spotlight:

From renumics import spotlightspotlight.show(df)

You can use Add View button in the inspector view and select bboxs_persons and filepath in the border view to display the corresponding border with the image:

Embed Rich data

To make the data structured, we can use image embedding of various basic models (ie: dense vector representation). To do this, you can use further dimensionality reduction techniques such as UMAP or t-SNE to apply Vision Transformer (ViT) embeddings of the entire image to the structuring of the dataset, thus providing a 2D similarity map of the image. Additionally, you can use the output of a pretrained object detector to structure your data by classifying it by the size or number of objects it contains. Since the COCO dataset already provides this information, we can use it directly.

Since Spotlight integrates support for the google/vit-base-patch16-224-in21k (ViT) model and UMAP , So when you create various embeds using file paths, it will be automatically applied:

spotlight.show(df, embed=["filepath"])

通过上述代码，Spotlight 将各种嵌入进行计算，并应用 UMAP 在相似性地图中显示结果。其中，不同的颜色代表了主要的类别。据此，您可以使用相似性地图来浏览数据：

预训练YOLOv8的结果

可用于快速识别物体的Ultralytics YOLOv8，是一套先进的物体检测模型。它专为快速图像处理而设计，适用于各种实时检测任务，特别是在被应用于大量数据时，用户无需浪费太多的等待时间。

为此，您可以首先加载预训练模型：

From ultralytics import YOLOdetection_model = YOLO("yolov8n.pt")

并执行各种检测：

detections = []forfilepathindf["filepath"].tolist()：detection = detection_model(filepath)[0]detections.append({ "yolo_bboxs"：[np.array(box.xyxyn.tolist())[0]forboxindetection.boxes]、 "yolo_conf_persons": np.mean([np.array(box.conf.tolist())[0]. forboxindetection.boxes ifdetection.names[int(box.cls)] =="person"]), np.mean(]), "yolo_bboxs_persons"：[np.array(box.xyxyn.tolist())[0] forboxindetection.boxes ifdetection.names[int(box.cls)] =="person],"yolo_categories": np.array([np.array(detection.names[int(box.cls)])forboxindetection.boxes], "yolo_categories": np.array(),})df_yolo = pd.DataFrame(detections)

在12gb的GeForce RTX 4070 Ti上，上述过程在不到20秒的时间内便可完成。接着，您可以将结果包含在DataFrame中，并使用Spotlight将其可视化。请参考如下代码：

df_merged = pd.concat([df, df_yolo], axis=1)spotlight.show(df_merged, embed=["filepath"])

下一步，Spotlight将再次计算各种嵌入，并应用UMAP到相似度图中显示结果。不过这一次，您可以为检测到的对象选择模型的置信度，并使用相似度图在置信度较低的集群中导航检索。毕竟，鉴于这些图像的模型是不确定的，因此它们通常有一定的相似度。

当然，上述简短的分析也表明了，此类模型在如下场景中会遇到系统性的问题：

• 由于列车体积庞大，站在车厢外的人显得非常渺小
• 对于巴士和其他大型车辆而言，车内的人员几乎看不到
• 有人站在飞机的外面
• 食物的特写图片上有人的手或手指

您可以判断这些问题是否真的会影响您的人员检测目标，如果是的话，则应考虑使用额外的训练数据，来增强数据集，以优化模型在这些特定场景中的性能。

小结

综上所述，预训练模型和 Spotlight 等工具的使用，可以让我们的对象检测可视化过程变得更加容易，进而增强数据科学的工作流程。您可以使用自己的数据去尝试和体验上述代码。

译者介绍

陈峻（Julian Chen），51CTO社区编辑，具有十多年的IT项目实施经验，善于对内外部资源与风险实施管控，专注传播网络与信息安全知识与经验。

原文标题：How to Explore and Visualize ML-Data for Object Detection in Images，作者：Markus Stoll

链接：https://itnext.io/how-to-explore-and-visualize-ml-data-for-object-detection-in-images-88e074f46361。

The above is the detailed content of How to explore and visualize ML data for object detection in images. For more information, please follow other related articles on the PHP Chinese website!