Search Results for "transformer"

...It’s widely used as an educational reference for people learning transformers in vision and as a lightweight baseline for research prototypes. The project encourages experimentation—swap optimizers, change augmentations, or plug the transformer backbone into downstream tasks.

VGGT is a transformer-based framework aimed at unifying classic visual geometry tasks—such as depth estimation, camera pose recovery, point tracking, and correspondence—under a single model. Rather than training separate networks per task, it shares an encoder and leverages geometric heads/decoders to infer structure and motion from images or short clips.

The Transformer architecture has improved the performance of deep learning models in domains such as Computer Vision and Natural Language Processing. Together with better performance come larger model sizes. This imposes challenges to the memory wall of the current accelerator hardware such as GPU. It is never ideal to train large models such as Vision Transformer, BERT, and GPT on a single GPU or a single machine.

Hiera is a hierarchical vision transformer designed to be fast, simple, and strong across image and video recognition tasks. The core idea is to use straightforward hierarchical attention with a minimal set of architectural “bells and whistles,” achieving competitive or superior accuracy while being markedly faster at inference and often faster to train. The repository provides installation options (from source or Torch Hub), a model zoo with pre-trained checkpoints, and code for evaluation and fine-tuning on standard benchmarks. ...

...The model takes sparse point queries on one frame and predicts their sub-pixel locations and a visibility score for every subsequent frame, producing long, coherent trajectories. Its transformer-style architecture aggregates information both along time and across points, allowing it to recover tracks even after brief disappearances. The repository ships with inference scripts, pretrained weights, and simple interfaces to seed points, run tracking, and export trajectories for downstream tasks. Typical uses include correspondence building, motion analysis, dynamic SLAM priors, video editing masks, and evaluation of geometric consistency in real scenes.

ConvNeXt is a modernized convolutional neural network (CNN) architecture designed to rival Vision Transformers (ViTs) in accuracy and scalability while retaining the simplicity and efficiency of CNNs. It revisits classic ResNet-style backbones through the lens of transformer design trends—large kernel sizes, inverted bottlenecks, layer normalization, and GELU activations—to bridge the performance gap between convolutions and attention-based models. ConvNeXt’s clean, hierarchical structure makes it efficient for both pretraining and fine-tuning across a wide range of visual recognition tasks. It achieves competitive or superior results on ImageNet and downstream datasets while being easier to deploy and train than transformers. ...

PyTorch training code and pretrained models for DETR (DEtection TRansformer). We replace the full complex hand-crafted object detection pipeline with a Transformer, and match Faster R-CNN with a ResNet-50, obtaining 42 AP on COCO using half the computation power (FLOPs) and the same number of parameters. Inference in 50 lines of PyTorch. What it is. Unlike traditional computer vision techniques, DETR approaches object detection as a direct set prediction problem.