SEEM, a universal segmentation model created by a Chinese team, takes one-time segmentation to a new level

Early this month, Meta released the “Segment Anything” AI model—Segment Anything Model (SAM). SAM is considered a universal basic model for image segmentation. It learns general concepts about objects and can generate masks for any object in any image or video, including objects and image types that have not been encountered during the training process. This "zero-sample migration" capability is amazing, and some even say that the CV field has ushered in a "GPT-3 moment."

Recently, a new paper "Segment Everything Everywhere All at Once" has once again attracted attention. In this paper, several Chinese researchers from the University of Wisconsin-Madison, Microsoft, and Hong Kong University of Science and Technology proposed a new prompt-based interaction model SEEM. SEEM can segment all content in an image or video at once and identify object categories based on various modal inputs given by the user (including text, images, graffiti, etc.). The project has been open sourced and a trial address is provided for everyone to experience.

Paper link: https://arxiv.org/pdf/2304.06718.pdf

Project link: https://github.com/UX-Decoder/Segment-Everything-Everywhere-All-At-Once

Trial address: https://huggingface.co/spaces/xdecoder/SEEM

This study verified the effectiveness of SEEM through comprehensive experiments Effectiveness on various segmentation tasks. Even though SEEM does not have the ability to understand user intentions, it shows strong generalization capabilities because it learns to write different types of prompts in a unified representation space. Furthermore, SEEM can efficiently handle multiple rounds of interactions through a lightweight prompt decoder.

Let’s take a look at the segmentation effect first:

In Transformers Segment "Optimus Prime" in the group photo:

It can also segment a type of object, such as segmenting a landscape picture. All buildings:

SEEM can also easily segment moving objects in the video:

This segmentation effect can be said to be very smooth. Let’s take a look at the approach proposed in this study.

Method Overview

This research aims to propose a general interface for image segmentation with the help of multi-modal prompts. In order to achieve this goal, they proposed a new scheme containing 4 attributes, including versatility, compositionality, interactivity and semantic-awareness, including

1) Versatility This research proposes to combine heterogeneous elements such as points, masks, text, detection boxes (boxes) and even the reference region of another image (referred region) , encoded into prompts in the same joint visual semantic space.

2) Compositionality writes queries on the fly for reasoning by learning a joint visual semantic space of visual and text prompts. SEEM can handle any combination of input prompts.

3) Interactivity: This study introduces the retention of conversation history information by combining learnable memory prompts and cross-attention guided by masks.

4) Semantic awareness: Use a text encoder to encode text queries and mask labels, thus providing open-set semantics for all output segmentation results.

Architecture-wise, SEEM follows a simple Transformer encoder-decoder architecture, and adds an additional text encoding device. In SEEM, the decoding process is similar to generative LLM, but with multimodal inputs and multimodal outputs. All queries are fed back to the decoder as prompts, and image and text encoders are used as prompt encoders to encode all types of queries.

Specifically, this study encodes all queries (such as points, boxes, and masks) into visual prompts, Also use a text encoder to convert text queries into text prompts so that visual and text prompts remain aligned. Five different types of prompts can all be mapped into the joint visual semantic space, and unseen user prompts can be processed through zero-shot adaptation. By training on different segmentation tasks, the model has the ability to handle various prompts. Additionally, different types of prompts can assist each other with cross-attention. Ultimately, SEEM models can use various prompts to achieve superior segmentation results.

In addition to its strong generalization capabilities, SEEM is also very efficient in operation. The researchers used prompts as input to the decoder, so SEEM only had to run the feature extractor once at the beginning over multiple rounds of interaction with humans. On each iteration, just run a lightweight decoder again with a new prompt. Therefore, when deploying the model, the feature extractor with large number of parameters and heavy running burden can be run on the server, while only the relatively lightweight decoder is run on the user's machine to alleviate the network latency problem in multiple remote calls. .

As shown in Figure 3(b) above, in multiple rounds of interaction, each interaction contains a manual loop and a model loop. In the artificial loop, the human receives the mask output of the previous iteration and gives positive or negative feedback for the next round of decoding through visual prompts. During the model loop, the model receives and updates memory prompts for future predictions.

Experimental results

This study experimentally compared the SEEM model with the SOTA interactive segmentation model, and the results are shown in Table 1 below.

As a general model, SEEM achieves comparable performance to models such as RITM, SimpleClick, and is comparable to the performance of SAM Very close, while SAM uses 50 times more segmented data for training than SEEM.

Unlike existing interactive models, SEEM is the first universal interface that supports not only classic segmentation tasks but also various user input types, including text, points, scribbles, Frames and images, providing powerful combination capabilities. As shown in Table 2 below, by adding combinable prompts, SEEM has significantly improved segmentation performance in cIoU, mIoU and other indicators.

Let’s take a look at the visual results of interactive image segmentation. Users only need to draw a point or simply scribble, and SEEM can provide very good segmentation results

You can also enter Text, let SEEM perform image segmentation

You can also directly input the reference image and point out the reference area, segment other images, and find objects that are consistent with the reference area:

This project is now available for online trial. Interested readers should go and try it.

The above is the detailed content of SEEM, a universal segmentation model created by a Chinese team, takes one-time segmentation to a new level. For more information, please follow other related articles on the PHP Chinese website!