Home

Awesome

Semantics Disentangling for Generalized Zero-Shot Learning

This is the official implementation for paper

Zhi Chen, Yadan Luo, Ruihong Qiu, Zi Huang, Jingjing Li, Zheng Zhang.
Semantics Disentangling for Generalized Zero-Shot Learning
International Conference on Computer Vision (ICCV) 2021.

Semantics Disentangling for Generalized Zero-shot Learning


Supplementary Experimental Results

In the paper, we followed the datasets provided in [15], in which the visual features in FLO dataset are normalized, and the semantic description of CUB dataset is the CNN-RNN sentence embeddings (1024D). We hereby provide extra comparison results on the visual features provided by GBU setting [17].

ModelAwA2 T1usHaPY T1usHCUB-EMB T1usHCUB-ATT T1usH
LFGAA [4]68.127.093.441.9--------67.636.280.950.0
DVBE [7]-63.670.867.0-32.658.341.8-----53.260.256.5
DVBE* [7]-62.777.569.4-37.955.945.2-----64.473.268.5
f-CLS WGAN[1]65.356.165.560.440.532.961.7--50.358.354.057.343.757.749.7
CANZSL[8]68.949.770.258.2--------60.647.958.152.5
LisGAN [9]70.652.676.362.343.134.368.245.7----58.846.557.951.6
CADA-VAE[10]64.055.875.063.9-31.755.140.3-52.054.853.461.851.653.552.4
f-VAEGAN-D2[11]71.157.670.663.5--------61.048.460.153.6
f-VAEGAN-D2*[11]70.357.176.165.2--------72.963.275.668.9
DLFZRL [12]70.3--60.946.7--38.5----61.8--51.9
TF-VAEGAN [13]72.259.875.166.6--------64.952.864.758.1
TF-VAEGAN*[13]73.455.583.666.7--------74.363.879.370.7
E-PGN [15]73.452.683.564.6----72.452.061.156.2----
AGZSL [16]73.865.178.971.341.035.165.545.7----57.241.449.745.2
AGZSL*[16]76.452.686.576.843.736.258.644.8----77.269.276.472.6
SDGZSL72.164.673.668.845.438.057.445.775.559.966.463.062.851.558.754.9
SDGZSL*74.369.686.573.747.036.260.747.578.573.077.575.173.766.075.970.6
ModelFLO-GBU T1usHFLO-EPGN T1usHSUN T1usH
LFGAA [4]--------61.518.540.025.3
DVBE [7]---------45.037.240.7
DVBE* [7]---------44.141.642.8
f-CLSWGAN[1]67.259.073.865.6----60.842.636.639.4
CANZSL[8]69.758.277.666.5----60.146.835.040.0
LisGAN [9]69.657.783.868.3----61.742.937.840.2
CADA-VAE[10]65.251.675.661.3----61.847.235.740.6
f-VAEGAN-D2[11]67.756.874.964.6----64.745.138.041.3
f-VAEGAN-D2*[11]70.463.392.475.1----65.650.137.843.1
DLFZRL [12]--------61.3--42.5
TF-VAEGAN [13]70.862.584.171.7----66.045.640.743.0
TF-VAEGAN*[13]74.763.892.579.4----66.741.851.946.3
E-PGN [15]----85.771.582.276.5----
AGZSL [16]----82.763.594.075.763.329.940.234.3
AGZSL*[16]----86.973.791.981.766.250.543.146.5
SDGZSL73.362.279.369.885.483.390.286.662.448.236.141.3
SDGZSL*76.673.288.780.286.986.189.187.865.251.140.245.0

TODO

Requirements

The implementation runs on

Usage

Put your datasets in SDGZSL_data folder and run the scripts in the folder.

The extracted features for APY and AWA datasets are from [1], FLO and CUB datasets are from [2]. For the fine-tuned features, AWA,FLO and CUB are from [3]. The APY fine-tuned features are extracted from us.

[1] Xian, Yongqin, et al. "Feature generating networks for zero-shot learning." CVPR 2018.

[2] Yu, Yunlong, et al. "Episode-based prototype generating network for zero-shot learning." CVPR 2020.

[3] Narayan, Sanath, et al. "Latent embedding feedback and discriminative features for zero-shot classification." ECCV 2020.

[4] Y. Liu, et al. "Attribute attention for semantic disambiguation in zero-shot learning." CVPR 2019.

[7] S. Min, et al. "Domain-aware visual bias eliminating for generalized zeroshot learning." CVPR 2020.

[8] Z. Chen, et al, "CANZSL: Cycleconsistent adversarial networks for zero-shot learning from natural language," WACV, 2020.

[9] J. Li, et al. "Leveraging the invariant side of generative zero-shot learning." CVPR, 2019.

[10] E. Schonfeld, et al. "Generalized zero-and few-shot learning via aligned variational autoencoders." CVPR, 2019.

[11] Y. Xian, et al. "f-vaegan-d2:A feature generating framework for any-shot learning." CVPR, 2019.

[12] B. Tong, et al. "Hierarchical disentanglement of discriminative latent features for zero-shot learning." CVPR, 2019.

[13] S. Narayan, et al. "Latent embedding feedback and discriminative features for zero-shot classification." ECCV, 2020.

[15] Y. Yu, et al. "Episode-based prototype generating network for zero-shot learning." CVPR, 2020.

[16] C., Yu-Ying, et al. "Adaptive and generative zero-shot learning." ICLR, 2020.

[17 ] Y. Xian, et al. "Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly." TPAMI, 2018.

Citation:

If you find this useful, please cite our work as follows:

@inproceedings{chen2021semantics,
	title={Semantics Disentangling for Generalized Zero-shot Learning},
	author={Chen, Zhi and Luo, Yadan and Qiu, Ruihong and Huang, Zi and Li, Jingjing and Zhang, Zheng},
	booktitle={ICCV},
	year={2021}
}