# GIFT: Learning Transformation-Invariant Dense Visual Descriptors via Group CNNs

Source: [GIFT: Learning Transformation-Invariant Dense Visual Descriptors via Group CNNs](<https://www.sota2.com/research/paper/68fa66af1e209131cea11c72>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-25T21:56:21.324Z

Authors: Yuan Liu, Zehong Shen, Zhixuan Lin, Sida Peng, Hujun Bao, Xiaowei Zhou

Published: 2019-11-14T04:20:57.000Z

## Abstract

Finding local correspondences between images with different viewpoints requires local descriptors that are robust against geometric transformations. An approach for transformation invariance is to integrate out the transformations by pooling the features extracted from transformed versions of an image. However, the feature pooling may sacrifice the distinctiveness of the resulting descriptors. In this paper, we introduce a novel visual descriptor named Group Invariant Feature Transform (GIFT), which is both discriminative and robust to geometric transformations. The key idea is that the features extracted from the transformed versions of an image can be viewed as a function defined on the group of the transformations. Instead of feature pooling, we use group convolutions to exploit underlying structures of the extracted features on the group, resulting in descriptors that are both discriminative and provably invariant to the group of transformations. Extensive experiments show that GIFT outperforms state-of-the-art methods on several benchmark datasets and practically improves the performance of relative pose estimation.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Camera Pose Estimation on MVS](<https://www.sota2.com/research/sota/camera-pose-estimation-on-mvs.md>): Camera pose estimation; MVS; mAA @ 20 deg: 60.

- [Local Feature Matching on HPatches illumination](<https://www.sota2.com/research/sota/local-feature-matching-on-hpatches-illumination.md>): Local Feature Matching; HPatches illumination; MMA@5px: 79.71.

- [Stereo Track on IMC Stereo track 2021 (val)](<https://www.sota2.com/research/sota/stereo-track-on-imc-stereo-track-2021-val.md>): Stereo track; IMC Stereo track 2021 (val); mAA (5°): 31.3.

- [Local Feature Matching on HPatches (all)](<https://www.sota2.com/research/sota/local-feature-matching-on-hpatches-all.md>): Local Feature Matching; HPatches (all); MMA@5px: 76.03.

- [Local Feature Matching on HPatches (viewpoint)](<https://www.sota2.com/research/sota/local-feature-matching-on-hpatches-viewpoint.md>): Local Feature Matching; HPatches (viewpoint); MMA (5px): 72.48.

- [Local Descriptor Matching on Roto-360 1.0 (test)](<https://www.sota2.com/research/sota/local-descriptor-matching-on-roto-360-1-0-test.md>): Local Descriptor Matching; Roto-360 1.0 (test); MMA @10px: 42.35.

- [Image Matching on ERDNIM (Day)](<https://www.sota2.com/research/sota/image-matching-on-erdnim-day.md>): Image Matching; ERDNIM (Day); HEstimation: 0.108.

- [Image Matching on ERDNIM (Night)](<https://www.sota2.com/research/sota/image-matching-on-erdnim-night.md>): Image Matching; ERDNIM (Night); HEstimation: 0.151.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/1911.05932>)

[PDF](<https://arxiv.org/pdf/1911.05932.pdf>)