Teaching a computer to find the exact object described in a sentence usually demands vast amounts of carefully labeled pictures. This paper shows how carefully made artificial examples can help when those pictures are scarce. The task is to locate a specific image region based on a free-form textual query: given a sentence, the system must find the matching part of a picture.

However, constructing these picture-and-description datasets with detailed region labels is both time-intensive and costly, which hinders progress. That creates the central challenge: learning effectively when only a few real training examples are accessible.

With scarce examples, the paper proposes POBF to address data scarcity; rather than generating whole pictures from scratch, it inpaints outside the box. It paints new surroundings outside an existing marked region, tackling label misalignment encountered in previous work by inpainting outside the box.

The approach then selects the most effective artificial examples instead of using everything it creates. In summary, the approach synthesizes high-quality training samples by filling in images and generating descriptions, then uses a filtering scheme for data selection.

So the idea has two parts: make additional picture-and-description pairs, then keep the ones most useful for learning. The process begins by synthesizing images and captions using image filling and captioning techniques, drawing on systems trained on large collections of image-and-description pairs.

A model trained on the limited real data then evaluates each artificial image, assigning scores that reflect its difficulty, its risk of encouraging overfitting, and a penalty term. Only the selected artificial images are retained and combined with the real data to train the final model.

Synthetic training images are useful only when the marked box still points to the right object. Here, the proposed approach keeps the cows and batters natural while placing the boxes accurately, unlike alternatives that distort the scene, miss the object, or shift the box.

The comparisons show that artificial data consistently outperforms training only on real data, and this approach gains an average of 5.83 percent over that baseline. POBF outperforms a baseline that directly generates training data, since GeoDiffusion can fail to generate an image for a novel target.

It also outperforms image-editing baselines: the strongest, X-Paste, averages thirty-one point three two percent, while POBF exceeds it by two point two nine percent. X-Paste produces unnatural images and Gen2Det has label misalignment, whereas POBF produces high-quality natural images with accurately aligned bounding boxes.

The improvement is more modest when the target objects are small, because filling very large background regions may introduce visible artifacts. Synthetic images improve performance by three point five three percent on average, demonstrating the effectiveness of the proposed inpainting strategy.

The results also separate the value of artificial descriptions: synthetic texts produce a more modest improvement of one point five one percent, showing their importance for performance gains. Adding synthetic texts produces a more modest improvement of 1.51 percent, showing that the generated descriptions still matter for performance gains.

The approach remains helpful when the available training data is limited: with only a small fraction of the original data, it improves accuracy from 13.97 percent to 19.18 percent. The consistent improvements across different amounts of training data highlight its potential for real-world applications where labeled data is limited.

Across the tested settings, the approach gains 5.83 percent over training only on real data and also surpasses leading alternatives, with accuracy gains ranging from 2.74 percent to 4.35 percent. Separate tests confirm that both the data-generation strategy and the filtering scheme contribute significant improvements.

The results remain robust across different image-making systems, training-data amounts, and model designs, and the authors present these findings as directions for future research in training-data generation and data selection. The key lesson is that more training data is not enough: the data must keep the picture and its description aligned, and the strongest examples must be chosen.

That could make image-search tools work with far less human labeling.