FSS-1000: A 1000-Class Dataset for Few-Shot Segmentation

By Tianhan Wei, Xiang Li, Yau Pun Chen, Yu-Wing Tai, Chi-Keung Tang. 2019

325 citations recorded by the source corpus.

Over the past few years, we have witnessed the success of deep learning in image recognition thanks to the availability of large-scale human-annotated datasets such as PASCAL VOC, ImageNet, and COCO. Although these datasets have covered a wide range of object categories, there are still a significant number of objects that are not included. Can we perform the same task without a lot of human annotations? In this paper, we are interested in few-shot object segmentation where the number of annotated training examples are limited to 5 only. To evaluate and validate the performance of our approach, we have built a few-shot segmentation dataset, FSS-1000, which consists of 1000 object classes with pixelwise annotation of ground-truth segmentation. Unique in FSS-1000, our dataset contains significant number of objects that have never been seen or annotated in previous datasets, such as tiny daily objects, merchandise, cartoon characters, logos, etc. We build our baseline model using standard backbone networks such as VGG-16, ResNet-101, and Inception. To our surprise, we found that training our model from scratch using FSS-1000 achieves comparable and even better results than training with weights pre-trained by ImageNet which is more than 100 times larger than FSS-1000. Both our approach and dataset are simple, effective, and easily extensible to learn segmentation of new object classes given very few annotated training examples. Dataset is available at https://github.com/HKUSTCV/FSS-1000

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Evaluation & Benchmarks, Meta & Continual Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Evaluation & Benchmarks

In Meta & Continual Learning

Citations in the corpus

Language-driven Semantic Segmentation (2022, methodology): “Hu et al. (2020) propose to use uncertainty-aware learning to better handle noisy labels of seen classes, while Li et al. (2020b) introduce a structured learning approach to better exploit the relations between seen and unseen categories.”
On the Binding Problem in Artificial Neural Networks (2020, background): “…physical reasoning (Chang et al., 2017; Battaglia et al., 2016; Janner et al., 2019), hierarchical physical reasoning (Mrowca et al., 2018; Li et al., 2020; Stanić et al., 2020), visual question answering (Santoro et al., 2017; Palm et al., 2018), abstract visual reasoning (Andreas,…”
Rethinking Space-Time Networks with Improved Memory Coverage for Efficient Video Object Segmentation (2021, methodology): “Following previous practices [18, 21], we ﬁrst pretrain the model on static image datasets [73, 74, 75, 76, 77] with synthetic deformation then perform main training on YouTubeVOS [78] and DAVIS [3, 65].”
OpenMask3D: Open-Vocabulary 3D Instance Segmentation (2023, background): “Given that the English language encompasses numerous nouns, totaling in the range of several hundred thousand [63], it is clear that existing closed-vocabulary approaches have an important limitation in terms of handling object categories and descriptions.”
Segment Anything in High Quality (2023, methodology): “HQSeg-44K is a collection of six existing image datasets including DIS [34] (train set), ThinObject-5K [29] (train set), FSS [26], ECSSD [37], MSRA-10K [8], DUT-OMRON [45] with extremely fine-grained mask labeling, where each of them contains 7.4K mask labels on average.”
OmniGen: Unified Image Generation (2024, unknown): “To the best of our knowledge, OmniGen is the first model capable of such flexible con-trol in image generation.”
Few-Shot Segmentation Without Meta-Learning: A Good Transductive Inference Is All You Need? (2020, background): “Few-shot segmentation, which has gained popularity recently [10, 17, 19, 23, 25, 33, 36, 37, 38, 39, 41, 42], is no exception.”
Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (2023, background): “The number of classes of a dataset is usually limited from tens [8] to hundreds [50], which is insignificant to English words [21].”
Personalize Segment Anything Model with One Shot (2023, methodology): “In Table 3, we evaluate our approach for one-shot image segmentation respectively on four datasets, FSS-1000 (Li et al., 2020), LVIS-92 i (Gupta et al., 2019), PASCAL-Part (Morabia et al., 2020), and PACO-Part (Ramanathan et al., 2023), where we follow Matcher (Liu et al., 2023) for data pre-process”
Prior Guided Feature Enrichment Network for Few-Shot Segmentation (2020, unknown): “FSS-1000 is composed of 1,000 classes, among which 486 classes are not included in any other existing datasets [17].”
Generalized Few-shot Semantic Segmentation (2020, background): “Few-Shot Segmentation Few-Shot Segmentation (FSSeg) places semantic segmentation in the few-shot scenario [31, 33, 28, 16, 45, 9, 58, 2, 46, 22, 37, 11, 40, 19, 57], where dense pixel labeling is performed on new classes with only a few support samples.”
Few-Shot Semantic Segmentation with Democratic Attention Networks (2020, methodology): “Following [38], the 1000 categories are divided into three splits for training, validation and testing.”
XMem: Long-Term Video Object Segmentation with an Atkinson-Shiffrin Memory Model (2022, methodology): “As in the implementation of [3,4], we use a mix of single object datasets [13,14,6,19,2].”