ImageBind Facebook or Meta AI model

Author

Chukwuemeka Okparaeke
July 23, 2023

I. Introduction

Artificial Intelligence (AI) has made significant advancements in various domains, ranging from computer vision to natural language processing. One of the key challenges in AI is understanding and effectively utilizing information from multiple modalities such as images, text, audio, and more. The ImageBind project, developed by Facebook AI Research (FAIR) and Meta AI, addresses this challenge by introducing a novel approach to multi-modal learning.

A. Overview of the ImageBind project

ImageBind is a cutting-edge project that aims to create a joint embedding space capable of binding together information from six different modalities. These modalities include images, text, audio, depth, thermal, and IMU data. By learning a shared embedding space, ImageBind enables the integration and fusion of diverse data types, facilitating seamless interactions and knowledge transfer between different modalities.

The project leverages state-of-the-art techniques in deep learning and representation learning to build a unified framework for multi-modal data analysis. It seeks to overcome the limitations of traditional AI models that often treat different modalities independently, failing to exploit the rich associations and complementary information available across diverse data sources.

B. Importance of multi-modal learning in AI applications

Multi-modal learning is of paramount importance in AI applications due to the inherent multi-sensory nature of real-world data. In many real-world scenarios, information is presented through a combination of visual, textual, and auditory cues. For instance, understanding a scene involves analyzing both visual elements (such as objects, shapes, and colors) and textual context (such as captions or descriptions).

By enabling multi-modal learning, models like ImageBind have the potential to significantly enhance AI capabilities in various domains. For instance, in computer vision, multi-modal learning can enable more robust and comprehensive object recognition, scene understanding, and image generation. In natural language processing, combining text and visual information can enhance tasks like image captioning, visual question answering, and sentiment analysis. Additionally, multi-modal learning can have wide-ranging applications in fields such as healthcare, robotics, autonomous vehicles, and more.

The ImageBind project pushes the boundaries of AI research and development, providing a powerful tool for harnessing the collective knowledge embedded in different modalities. By unlocking the potential of multi-modal learning, ImageBind opens doors to new possibilities and applications, contributing to the advancement of AI technology.

In the following sections, we will delve deeper into the workings of ImageBind, explore its applications, discuss technical details and implementation, examine performance and results, and outline future developments and impact. Stay tuned to unravel the fascinating world of multi-modal learning with ImageBind.

II. Understanding ImageBind

ImageBind is a groundbreaking project that introduces a novel concept of a joint embedding space, enabling the integration and binding of information from diverse modalities. This section aims to provide a deeper understanding of ImageBind by explaining the joint embedding space concept, highlighting the six modalities it covers, and discussing its key features and capabilities.

A. Explanation of the joint embedding space concept

In ImageBind, a joint embedding space refers to a unified representation where data from different modalities are mapped to a shared feature space. This joint space allows for direct comparisons and interactions between different modalities, promoting seamless knowledge transfer and cross-modal analysis. By learning a joint embedding space, ImageBind facilitates the alignment and fusion of information from multiple sources, enabling holistic data understanding and integration.

The joint embedding space concept in ImageBind is based on the idea that different modalities often provide complementary information. For example, an image and its corresponding textual description can capture different aspects of a scene, with the image conveying visual details and the text providing contextual information. By embedding both the image and text into a shared space, ImageBind can measure their similarity or perform operations that combine their representations.

B. Six modalities covered: images, text, audio, depth, thermal, and IMU data

ImageBind stands out for its comprehensive coverage of six different modalities, encompassing a wide range of data sources. These modalities include:

  1. Images: Visual information captured in the form of images, such as photographs or frames from videos.

  2. Text: Linguistic data, including textual descriptions, captions, or any other form of written language.

  3. Audio: Sound data, such as speech, music, or environmental sounds.

  4. Depth: Depth information derived from depth sensors or techniques, providing spatial understanding and 3D structure.

  5. Thermal: Thermal imaging data that captures heat signatures and temperature distributions.

  6. IMU (Inertial Measurement Unit) data: Data from inertial sensors, including accelerometers, gyroscopes, and magnetometers, capturing motion and orientation.

By incorporating these diverse modalities, ImageBind aims to create a holistic understanding of data by leveraging multiple sources of information. This enables the model to exploit the unique characteristics and nuances present in each modality, leading to more robust and comprehensive analysis.

C. Key features and capabilities of ImageBind

ImageBind possesses several key features and capabilities that make it a powerful tool for multi-modal learning:

  1. Cross-modal retrieval: ImageBind enables searching and matching across different modalities. For example, it can retrieve relevant images based on a given text query or find similar text descriptions given an input image.

  2. Composing modalities with arithmetic: By operating in the joint embedding space, ImageBind allows for arithmetic operations on different modalities. For instance, it can generate novel image-text pairs by combining embeddings through addition or subtraction.

  3. Cross-modal detection and generation: ImageBind can detect and generate content across modalities. For example, it can detect objects in an image and generate textual descriptions or generate realistic images based on textual prompts.

These features and capabilities highlight the versatility and potential of ImageBind in diverse applications, including content-based retrieval, data synthesis, and multimodal understanding.

In the next section, we will explore the technical details and implementation of ImageBind, providing insights into how it can be

III. Applications of ImageBind

ImageBind, with its ability to create a joint embedding space for multiple modalities, opens up a wide range of applications in the field of AI. In this section, we will explore some of the key applications of ImageBind, including cross-modal retrieval, composing modalities with arithmetic, and cross-modal detection and generation.

A. Cross-modal retrieval: How ImageBind enables searching and matching across different modalities

One of the primary applications of ImageBind is cross-modal retrieval. Traditional information retrieval systems typically operate within a single modality, such as retrieving images based on visual similarity or finding text documents based on textual keywords. However, with ImageBind, cross-modal retrieval becomes possible, allowing users to search and match information across different modalities.

For example, given a text query, ImageBind can retrieve relevant images that are semantically related to the query. Similarly, given an input image, ImageBind can find text descriptions or documents that describe the content of the image. This cross-modal retrieval capability facilitates more comprehensive and intuitive searches, bridging the gap between different types of data and enhancing the user experience.

B. Composing modalities with arithmetic: Creating new data samples by combining modalities

ImageBind also enables the composition of modalities using arithmetic operations in the joint embedding space. This means that it becomes possible to combine and manipulate representations from different modalities to create new data samples.

For instance, by adding the embedding of a text description to the embedding of an image, ImageBind can generate a new embedding that represents the combination of both modalities. This allows for the creation of novel image-text pairs that possess characteristics from both sources. Similarly, subtraction operations can be used to remove specific modalities or attributes, resulting in modified data samples.

The ability to compose modalities through arithmetic operations opens up possibilities for data augmentation, creative synthesis, and enhanced data understanding. It can be particularly valuable in tasks such as image captioning, where generating diverse and contextually rich descriptions is crucial.

C. Cross-modal detection and generation: Generating data in one modality based on information from another

ImageBind's joint embedding space also enables cross-modal detection and generation. This means that given data in one modality, ImageBind can generate or detect corresponding information in another modality.

For example, given an image, ImageBind can detect objects or regions of interest within the image and generate textual descriptions that describe the detected objects. Similarly, given a text prompt, ImageBind can generate realistic images that align with the textual description.

This cross-modal detection and generation capability have significant implications for tasks such as image synthesis, text-to-image translation, and multimodal content generation. It allows for the creation of rich, diverse, and contextually aligned data in different modalities, enabling advancements in areas like computer vision, natural language processing, and multimedia content creation.

These applications highlight the versatility and potential of ImageBind in leveraging the power of multi-modal learning. By enabling cross-modal retrieval, composing modalities with arithmetic, and facilitating cross-modal detection and generation, ImageBind unlocks new possibilities for data analysis, synthesis, and understanding.

In the next section, we will delve into the technical details and implementation aspects of ImageBind, providing insights into how to utilize this powerful framework effectively

IV. Technical Details and Implementation

In this section, we will explore the technical details and implementation aspects of ImageBind. We will discuss the PyTorch implementation of the ImageBind model, the availability of pretrained models, and provide a step-by-step guide on how to extract and compare features across different modalities using ImageBind.

A. PyTorch implementation of the ImageBind model

ImageBind is implemented using the PyTorch framework, a popular choice for deep learning research and development. The PyTorch implementation of ImageBind provides a flexible and efficient platform for training, evaluating, and utilizing the model.

The source code for the ImageBind project can be found on the official GitHub repository: https://github.com/facebookresearch/ImageBind. It includes the necessary code files and dependencies to run ImageBind on your local machine or in a distributed computing environment.

B. Pretrained models and their availability

ImageBind also provides pretrained models, which are already trained on large-scale datasets and can be readily used for various applications without the need for extensive training. These pretrained models capture the learned joint embedding space and the knowledge of multiple modalities.

The availability of pretrained models depends on the specific version and release of ImageBind. It is recommended to refer to the GitHub repository or the official project documentation for the most up-to-date information on the availability and usage of pretrained models.

C. Extracting and comparing features across modalities using ImageBind

To leverage the capabilities of ImageBind for extracting and comparing features across different modalities, we can follow a series of steps. Here, we provide a code example along with an explanation of each step:

  1. Code example and explanation:

import data
import torch
from models import imagebind_model
from models.imagebind_model import ModalityType

text_list = ["A dog.", "A car", "A bird"]
image_paths = [".assets/dog_image.jpg", ".assets/car_image.jpg", ".assets/bird_image.jpg"]
audio_paths = [".assets/dog_audio.wav", ".assets/car_audio.wav", ".assets/bird_audio.wav"]

device = "cuda:0" if torch.cuda.is_available() else "cpu"

# Instantiate model
model = imagebind_model.imagebind_huge(pretrained=True)
model.eval()
model.to(device)

# Load data
inputs = {
    ModalityType.TEXT: data.load_and_transform_text(text_list, device),
    ModalityType.VISION: data.load_and_transform_vision_data(image_paths, device),
    ModalityType.AUDIO: data.load_and_transform_audio_data(audio_paths, device),
}

with torch.no_grad():
    embeddings = model(inputs)

print(
    "Vision x Text: ",
    torch.softmax(embeddings[ModalityType.VISION] @ embeddings[ModalityType.TEXT].T, dim=-1),
)
print(
    "Audio x Text: ",
    torch.softmax(embeddings[ModalityType.AUDIO] @ embeddings[ModalityType.TEXT].T, dim=-1),
)
print(
    "Vision x Audio: ",
    torch.softmax(embeddings[ModalityType.VISION] @ embeddings[ModalityType.AUDIO].T, dim=-1),
)

Explanation:

  • The code snippet imports the necessary modules and libraries, including the ImageBind model and the data module for loading and transforming data across modalities.

  • It defines lists of text, image, and audio inputs that we want to process and compare.

  • The device is set to GPU (cuda:0) if available; otherwise, it falls back to CPU.

  • The ImageBind model (imagebind_huge) is instantiated with pretrained weights and moved to the specified device.

  • The data inputs are loaded and transformed using the load_and_transform_text, load_and_transform_vision_data, and load_and_transform_audio_data functions from the data module. These functions handle the preprocessing and transformation of the input data to ensure compatibility with the ImageBind model.

With the data inputs prepared, we proceed to calculate the embeddings using the ImageBind model. The inputs are passed to the model using the model() function, and the resulting embeddings are stored in the embeddings variable.

Finally, we print the output of the cross-modal comparisons. The torch.softmax function is applied to the dot products of the embeddings for different modalities to obtain a probability distribution over the possible matches. This allows us to assess the similarity or compatibility between different modalities.

The expected output is shown as comments in the code snippet, providing an example of the results for the vision-text, audio-text, and vision-audio comparisons.

This code example demonstrates the ease with which ImageBind can be implemented and used to compare and combine modalities in a multi-modal learning setting. By leveraging the joint embedding space learned by ImageBind, researchers and practitioners can unlock the potential of multi-modal data and explore various applications across domains.

Note that this code snippet is a simplified example, and actual usage may require additional preprocessing steps or modifications based on specific data and model requirements. Nonetheless, it serves as a starting point for utilizing ImageBind and harnessing its

  1. Expected output:

The code snippet provided above performs feature extraction and comparison across modalities using ImageBind. Here's the expected output: 

Vision x Text:
tensor([[9.9761e-01, 2.3694e-03, 1.8612e-05],
        [3.3836e-05, 9.9994e-01, 2.4118e-05],
        [4.7997e-05, 1.3496e-02, 9.8646e-01]])

Audio x Text:
tensor([[1., 0., 0.],
        [0., 1., 0.],
        [0., 0., 1.]])

Vision x Audio:
tensor([[0.8070, 0.1088, 0.0842],
        [0.1036, 0.7884, 0.1079],
        [0.0018, 0.0022, 0.9960]])

The output shows the similarity scores between different modalities:

  • The "Vision x Text" section displays the softmax similarity scores between visual embeddings and text embeddings. Each row corresponds to a visual input (e.g., an image), and each column corresponds to a text input. Higher scores indicate higher similarity.

  • The "Audio x Text" section shows the softmax similarity scores between audio embeddings and text embeddings. Again, each row corresponds to an audio input, and each column corresponds to a text input.

  • The "Vision x Audio" section presents the softmax similarity scores between visual embeddings and audio embeddings.

These similarity scores provide insights into the relationships between different modalities in the joint embedding space, indicating how similar or related the representations are across modalities.

By following this example, users can leverage ImageBind to extract and compare features from various modalities, facilitating tasks such as cross-modal retrieval, content matching, and modality-specific analysis.

In the next section, we will discuss the model card, licensing, contributing guidelines, and how to cite ImageBind for academic or research purposes.

V. Performance and Results

In this section, we will explore the performance and results achieved by ImageBind. We will discuss the evaluation metrics used to assess its performance, highlight its zero-shot classification performance across different datasets, and compare it with other approaches or models in the field.

A. Evaluation metrics for ImageBind

To evaluate the performance of ImageBind, various metrics are employed to measure its effectiveness in different tasks. These metrics depend on the specific application and the dataset being used. Some common evaluation metrics include accuracy, precision, recall, F1 score, and mean average precision (mAP). These metrics provide quantitative measures of the model's performance in tasks such as cross-modal retrieval, classification, and generation.

B. Zero-shot classification performance across different datasets

ImageBind has demonstrated impressive zero-shot classification performance across various datasets. Zero-shot classification refers to the ability to classify samples from unseen classes, without any explicit training on those classes. ImageBind achieves this by learning a joint embedding space that captures the relationships between different modalities.

For example, ImageBind has achieved the following zero-shot classification performance on different datasets:

  • IN1k: 77.7% accuracy

  • K400: 50.0% accuracy

  • NYU-D: 54.0% accuracy

  • ESC: 66.9% accuracy

  • LLVIP: 63.4% accuracy

  • Ego4D: 25.0% accuracy

These results highlight the effectiveness of ImageBind in capturing the underlying semantic relationships between modalities and its ability to generalize to unseen classes.

C. Comparison with other approaches or models

ImageBind has been compared with other approaches or models in the field of multi-modal learning. Through these comparisons, ImageBind has demonstrated its superiority in various tasks and datasets. It has showcased its effectiveness in capturing and leveraging the joint embedding space for enhanced performance.

In comparison to other models, ImageBind has exhibited competitive performance in tasks such as cross-modal retrieval, modality composition, and cross-modal detection and generation. Its ability to seamlessly integrate information from multiple modalities and leverage the joint embedding space sets it apart from traditional approaches.

However, it is important to note that the performance comparison may vary based on the specific dataset, task, and evaluation metrics employed. Researchers and practitioners are encouraged to conduct their own evaluations and comparisons based on their specific requirements.

Overall, ImageBind has shown promising results and outperformed other approaches in several multi-modal learning tasks, making it a valuable tool for various AI applications.

In the next section, we will discuss the model's limitations, ongoing research,

VI. Model Card and Licensing

In this section, we will explore the model card for ImageBind, providing important details about the model's characteristics, limitations, intended usage, and potential biases. Additionally, we will discuss the licensing information for the ImageBind code and model weights.

A. Details on the ImageBind model card

The ImageBind project includes a comprehensive model card that provides crucial information about the model. The model card serves as a documentation resource, ensuring transparency and promoting responsible AI usage. It contains details such as:

  • Model characteristics: The model card outlines the architecture, design choices, and implementation details of ImageBind. It describes how the joint embedding space is learned across six different modalities: images, text, audio, depth, thermal, and IMU data.

  • Intended usage: The model card provides guidance on the appropriate and recommended usage of ImageBind. It specifies the tasks for which ImageBind is well-suited, along with any limitations or considerations to keep in mind when applying the model.

  • Limitations and biases: The model card addresses the limitations of ImageBind, including potential biases and areas where the model may perform suboptimally. This information is crucial for users to understand the model's scope and limitations in order to make informed decisions.

The model card is an important resource for researchers, practitioners, and developers who utilize ImageBind in their projects. It promotes ethical and responsible AI practices by providing transparency and accountability.

B. License information for ImageBind code and model weights

The ImageBind code and model weights are released under the CC-BY-NC 4.0 license. This license allows users to freely use, modify, and distribute the code and weights for non-commercial purposes. It ensures that the ImageBind project remains open and accessible to the research community, enabling collaboration and further advancements in multi-modal learning.

Users should refer to the LICENSE file for additional details on the licensing terms and conditions associated with ImageBind. It is important to adhere to the license and respect the copyright and intellectual property rights of the ImageBind project.

By providing the code and model weights under an open-source license, the creators of ImageBind encourage collaboration, innovation, and the development of new applications and research in the field of multi-modal learning.

In the final section, we will conclude the article by summarizing the key points discussed and highlighting the significance of ImageBind in the field of AI.

VII. Future Developments and Impact

In this section, we will explore the potential advancements and extensions for ImageBind, as well as its impact on various industries through real-world applications.

A. Potential advancements and extensions for ImageBind

ImageBind has laid a solid foundation for multi-modal learning, but there are several potential advancements and extensions that can further enhance its capabilities:

  1. Expansion to additional modalities: While ImageBind already covers six modalities, there is room for incorporating additional modalities such as video, motion, or physiological signals. Expanding the joint embedding space to encompass more diverse data types can unlock new possibilities and applications.

  2. Improved scalability: As the size and complexity of multi-modal datasets continue to grow, advancements in scalability are crucial. Further research can focus on developing efficient algorithms and architectures to handle large-scale multi-modal datasets and improve the scalability of ImageBind.

  3. Continual learning and adaptation: Enabling ImageBind to continuously learn and adapt to new data or modalities can enhance its flexibility and applicability. Techniques such as online learning and lifelong learning can be explored to enable incremental updates to the joint embedding space.

  4. Interpretability and explainability: While ImageBind effectively learns the joint embedding space, understanding the learned representations and providing interpretability and explainability is an important research direction. Enhancing the interpretability of ImageBind can enable users to gain insights into the underlying relationships between modalities and improve trust in the model's decisions.

B. Real-world applications and impact on various industries

ImageBind has the potential to revolutionize various industries and domains through its multi-modal learning capabilities. Some notable real-world applications and their potential impact include:

  1. Healthcare: ImageBind can facilitate more accurate and comprehensive medical diagnoses by combining multiple modalities such as medical images, patient records, and sensor data. It can aid in disease detection, treatment planning, and personalized healthcare delivery.

  2. Media and Entertainment: By leveraging ImageBind, content creators can generate engaging and immersive experiences by combining visual, audio, and textual elements. It can enable the creation of interactive media, virtual reality applications, and content recommendation systems.

  3. Autonomous Systems: ImageBind can play a vital role in autonomous systems, such as self-driving cars and robots. By integrating and interpreting data from various sensors and modalities, it can enhance perception, decision-making, and situational awareness in real-time.

  4. E-commerce and Recommendation Systems: ImageBind's cross-modal retrieval capabilities can revolutionize e-commerce platforms by enabling users to search and match products based on diverse modalities. It can improve recommendation systems by considering user preferences across multiple dimensions.

These are just a few examples of how ImageBind can impact different industries. Its ability to bind multiple modalities in a joint embedding space opens up possibilities for innovative applications and advancements across various domains.

In conclusion, ImageBind represents a significant breakthrough in multi-modal learning. With its joint embedding space, it enables cross-modal retrieval, composition, detection, and generation, leading to enhanced performance and new possibilities in AI applications. As research and development continue, ImageBind is expected to have a profound impact on industries, fostering innovation and advancements in multi-modal AI systems.

VIII. Conclusion

In this article, we have explored ImageBind, a groundbreaking project that introduces the concept of a joint embedding space to facilitate multi-modal learning in AI. Let's summarize the key points discussed and emphasize the importance of ImageBind in advancing multi-modal learning.

A. Summary of the key points discussed

  1. Overview of the ImageBind project: ImageBind aims to learn a joint embedding across six modalities - images, text, audio, depth, thermal, and IMU data. It enables cross-modal retrieval, composition, detection, and generation, offering a wide range of applications in AI.

  2. Importance of multi-modal learning in AI applications: Multi-modal learning allows AI models to leverage information from different modalities, leading to a deeper understanding of data. By integrating multiple modalities, models like ImageBind can capture richer and more comprehensive representations, enhancing performance and enabling novel applications.

  3. Explanation of the joint embedding space concept: ImageBind learns a shared embedding space where different modalities can be aligned and compared. This joint embedding allows for seamless integration and interaction between modalities, enabling cross-modal operations and facilitating multi-modal tasks.

  4. Six modalities covered by ImageBind: ImageBind covers a diverse set of modalities, including images, text, audio, depth, thermal, and IMU data. This broad coverage enables the model to capture and leverage information from various sources, leading to a holistic understanding of the data.

  5. Applications of ImageBind: ImageBind has several practical applications, including cross-modal retrieval, composing modalities with arithmetic, and cross-modal detection and generation. These applications have significant implications in fields such as healthcare, media, autonomous systems, and e-commerce.

  6. Technical details and implementation: We explored the PyTorch implementation of ImageBind, including the extraction and comparison of features across modalities. A code example demonstrated how to utilize ImageBind for multi-modal tasks, and the expected output showcased the model's capabilities.

  7. Performance and results: ImageBind demonstrates impressive zero-shot classification performance across different datasets. Its effectiveness in learning joint representations outperforms previous approaches, showcasing its potential in multi-modal learning tasks.

  8. Model card and licensing: ImageBind provides a model card that outlines important details about the model, including its characteristics, limitations, and intended usage. The ImageBind code and model weights are released under the CC-BY-NC 4.0 license, ensuring open access and fostering collaboration in the research community.

B. Importance of ImageBind in advancing multi-modal learning in AI

ImageBind represents a significant advancement in the field of multi-modal learning and has several implications for AI research and applications. By enabling joint embedding across multiple modalities, ImageBind breaks down the barriers between different data types and facilitates seamless integration and interaction. This leads to a deeper understanding of data and enables AI models to leverage diverse sources of information.

The importance of ImageBind in advancing multi-modal learning lies in its ability to unlock new possibilities and applications. By leveraging multiple modalities, models like ImageBind can address real-world challenges that require a holistic understanding of data. Whether it's in healthcare, media, autonomous systems, or e-commerce, the integration of multiple modalities through ImageBind enhances performance, enables innovative applications, and drives progress in AI.

With ongoing research and future developments, ImageBind is poised to have a lasting impact on the field of multi-modal learning. Its potential for expansion to additional modalities, scalability improvements, and interpretability enhancements will further enhance its capabilities and widen its range of applications.

In conclusion, ImageBind represents a significant advancement in multi-modal learning, bridging the gap between different data modalities and enabling powerful interactions. Its ability to learn joint embeddings across modalities unlocks new possibilities and facilitates seamless integration and interaction between diverse data types. The applications of ImageBind are far-reaching and have the potential to revolutionize various industries.

In fields such as healthcare, ImageBind can enable cross-modal retrieval, allowing medical professionals to search and match patient data across different modalities. This can lead to more accurate diagnoses and personalized treatment plans. Additionally, ImageBind's ability to compose modalities with arithmetic opens up possibilities for data augmentation and synthesis, providing researchers with new ways to generate diverse and representative training samples.

Media and entertainment industries can also benefit from ImageBind. Content creators can leverage the cross-modal detection and generation capabilities to automatically generate captions or descriptions for images or videos based on the available audio or text information. This can streamline content production processes and enhance user experiences by making media more accessible and engaging.

In the autonomous systems domain, ImageBind can contribute to the development of advanced perception systems. By integrating data from sensors such as depth, thermal, and IMU, ImageBind can enhance the perception capabilities of autonomous vehicles or robots. This can improve their ability to navigate complex environments, detect obstacles, and make informed decisions based on multiple modalities of information.

Furthermore, ImageBind has implications for the e-commerce industry. With ImageBind's cross-modal retrieval capabilities, online retailers can offer more intuitive and personalized product recommendations to their customers. By understanding the relationships between visual features, textual descriptions, and user preferences, ImageBind can facilitate more accurate matching between products and customer needs, leading to enhanced user satisfaction and increased sales.