This content originally appeared on HackerNoon and was authored by The FewShot Prompting Publication
:::info Authors:
(1) Yingxu He, Department of Computer Science National University of Singapore {e0139128@u.nus.edu};
(2) Qiqi Sun, College of Life Sciences Nankai University {sunqiqi1018@gmail.com}.
:::
Table of Links
2. Methodology
In this section, we describe our proposed approach to automatically collect captions for remote sensing images by guiding LLMs to describe their object annotations. In this work, we limit the number of objects in each image to no more than 15, which ensures a relatively simple spatial layout for the LLM. Our approach consists of three main steps: (1) develop APIs to conduct geographical analysis and describe spatial relationships between objects, (2) prompt the API to generate captions with the help from APIs, and (3) caption evaluation and selection. We explain each step in detail below.
2.1 Spatial Relationship APIs
LLM is incompetent at processing 2-dimensional geographical information, so we implemented several analytical approaches to analyze the spatial relations between objects. Inspired by the captions provided by the RSICD paper, we only focused on analyzing the distances between objects, the concentration of object locations, shapes formed by groups of objects, and significant relations between objects.
2.1.1 Distance
In the Xview and Dota datasets, the size of objects varies a lot. Therefore, using the distance between centers is inappropriate for the distances between objects. For instance, although the centers of two large buildings might be quite far apart, their inner-facing walls might be only a few steps away. Therefore, we consider the shortest distances between bounding boxes as their distance. For the distance between two groups of objects, we represent it with the distance between their closest element, which is normally referred to as the Single Linkage measure in the field of clustering.
2.1.2 Clustering
One of the most important features captured by human eyes is the concentration of objects based on their locations and types, e.g., one tends to easily differentiate a vehicle running on a highway from several buildings standing by the road. On the other hand, people also tend to pay attention to the objects’ closest neighbor, e.g., a passenger car next to a truck is easier to draw people’s attention than a building relatively further away from the truck. Traditional machine learning clustering algorithms include distance-based algorithms such as K-Means and hierarchical clustering, and density-based clustering such as DBSCAN and its variants. However, the K-Means algorithm often fails to separate outliers from concentrated objects, while the benefits of density-based clustering might be buried in this case, where each image only contains fewer than ten objects.
\ In this work, We used the Minimum Spanning Tree (MST) algorithm to connect all the objects in the image and form clusters by removing significantly long edges from the graph. Kruskal’s MST algorithm[3] considers objects’ nearest neighbors and simultaneously skips negligible connections, ensuring every tree edge is aligned to humans’ observing behavior. We set the threshold at the 75 percentile of the edge weights from the entire dataset. Edges above this threshold were removed from the graph to form clusters, minimizing intra-cluster and maximizing inter-cluster distances. To encourage grouping objects of the same type into the same cluster, We add extra length to distances between objects of different types. Figure 1 gives a detailed illustration on the MST-based clustering algorithm. This approach could precisely split objects by type, location, and proximity, which benefits the subsequent geographical analysis.
\
2.1.3 Geometric Shape
Inspired by the captions provided in the RSICD dataset, the line shape is considered the fundamental shape to be detected in this work. It seems most attractive to human eyes and the basic element of many other complicated shapes. For instance, the square grid street pattern is one of the most popular street patterns used in cities, where lines of buildings are the most fundamental elements. Undeniably, other shapes could also easily draw people’s attention, such as circles and squares. Nonetheless, in the setting of this work, where each image contains at most 15 objects, they are less obvious and more difficult to detect. Therefore, we only implemented a method to detect line shapes from groups of objects by inspecting whether the lines formed by corners of bounding boxes are parallel.
2.1.4 Geometric Relation
We review some relations listed in the RSICD paper[6] and come out with our list of relations to be included in the image captions: "stands alone", "near", "in a row", "surrounded by", "between", and "in two sides of". We modified the "in rows" relation from RSICD paper to "in a row", as objects in different rows can be clustered into different groups as is described in section 2.1.2, and any possible line shape will be detected by the shape identification algorithm described in section 2.1.3. Additionally, we propose a "between" relation as the flip side of "in two sides of" to differentiate the situation where there are only objects on the two sides of others from objects circling others 360◦ . In this work, the approaches described above can address relations "stands alone", "near", and "in a row". The relation "surrounded by" is only considered when certain objects are located within the border of another group of objects. The detailed function is achieved by drawing links from the boxes in the middle to the outer ones and calculating the angles between them. The implementation of relations "between" and "in two sides of" are left for future work.
2.2 LLM Prompting
The second step of our approach is to use prompts to guide the LLM to produce a caption following a similar pattern. With the APIs implemented in section 2.1, there are many options to prompt the LLM and guide it to generate the ideal captions. Following the recently popular idea of treating the LLMs as a controller or action dispatcher[13], one approach could be allowing the language model to plan its actions and execute the functions in sequences to obtain helpful geographical analysis results. For instance, the recently developed ReAct[10] approach synergizes the reasoning and executing process of LLM to enhance its capability of handling complex tasks. It allows great flexibility in geographical analysis and greater diversity in the generated captions. Nonetheless, the LLM tends to experience difficulty discovering eye-catching geographical relations and is easily flooded with less important information received during the action execution process.
\ To solve the problem, we adopted the advantage of the MST algorithm, which reveals the most important neighbors for both clusters and stand-alone objects, from where we can easily extract the significant geographical relations. More specifically, we list the presence of every group in each image with their combination and shapes detected, together with stand-alone objects. The significant geometric relations between the boxes are then provided to give the LLM a sense of their spatial relations. In this case, we only present the edges removed during the clustering step (section 2.1.2) that connects clusters and stand-alone objects. An illustration of the spatial relations presented and captions created by LLM is provided in figure 2.
2.2.1 Captions Diversification
Although the prompt already provided necessary clustering information and spatial relations between objects, LLM is not supposed only to bring the clustering information into the spatial relations and create captions, which can be already done by a template-based or rule-based method. The most important role played by LLM is to understand the current spatial layout and paraphrase the potentially redundant or insignificant relations into appropriate captions. For instance, in figure 2 (2), the MST-based algorithm detects one building is closer to some buildings than others. However, as the whole image is occupied with different buildings, a caption repeating that relation might bring confusion and ambiguity to the downstream deep-learning models and even human readers. In this case, LLM plays a vital role in evaluating the significance of each spatial relation and performing necessary paraphrasing.
\ In this work, the summarising behavior of LLM is ensured by providing necessary examples in the prompt, which is more frequently referred to as the "Few-Shot" prompting technique. We provided several examples where LLM is supposed to synergize the clustering results with the spatial relations to create captions in its own words. Other prompting techniques could potentially achieve the same goal, such as adding descriptions for the expected behaviors or breaking down the reasoning process using Chain of Thought or Tree of Thought techniques. Nonetheless, given the input and expected output format are already complicated, these prompting strategies could bring much more complexity and difficulty into the prompt writing process. Moreover, our experimental results show that few-shot prompting performs more stably than any of the above-mentioned techniques.
2.2.2 Response Formatting
Additionally, to effectively restrict the response to a computer-readable format, we explicitly instruct the LLM to output the captions in the format of a Python list, whose detailed information has already been included in LLM’s pre-training corpus and well embedded in its parametric memory, rather
\
\ than other customized format that requires extra explanation. It is desired not to have any id of the object groups in the LLM response, which is achieved again by providing examples in the prompt, as introduced in the prior section. It has been stated in many recent research works that few-shot prompting works better than zero-show prompting with prolonged instructions[11]. The detailed procedures can be shown in figure 3.
2.3 Caption Evaluation and Selection
The third step of our approach is to evaluate and select the best caption for each image. We use two criteria to assess the quality of captions: (a) caption quality, which measures how well the caption matches the ground truth annotation, and (b) caption diversity, which measures how different the caption is from other captions generated from other images. We use the following procedure:
\ • We filter out captions containing undesirable keywords such as the group’s id, like "group 0" or the group’s order, like "the first group", which could lead to confusion.
\ • We use pre-trained CLIP to compute a score for each caption based on its match the input image. The evaluator is trained on a large-scale image-caption dataset that covers various domains and scenarios.
\ • We use a similarity measure to compute a score for each caption based on caption diversity. The similarity measure compares each caption with captions generated from other images to avoid descriptions that are too vague and broad.
\ • We combine both scores using a weighted average formula to obtain a final score for each caption.
\ • We select the caption with the highest final score as the best caption for each image.
\
\ \
:::info This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license.
:::
\
This content originally appeared on HackerNoon and was authored by The FewShot Prompting Publication
The FewShot Prompting Publication | Sciencx (2024-06-16T14:30:21+00:00) Towards Automatic Satellite Images Captions Generation Using LLMs: Methodology. Retrieved from https://www.scien.cx/2024/06/16/towards-automatic-satellite-images-captions-generation-using-llms-methodology/
Please log in to upload a file.
There are no updates yet.
Click the Upload button above to add an update.