Area-aware pre-training for open-vocabulary object detection with imaginative and prescient transformers – Google Analysis Weblog

The power to detect objects within the visible world is essential for pc imaginative and prescient and machine intelligence, enabling functions like adaptive autonomous brokers and versatile purchasing programs. Nevertheless, trendy object detectors are restricted by the handbook annotations of their coaching knowledge, leading to a vocabulary dimension considerably smaller than the huge array of objects encountered in actuality. To beat this, the open-vocabulary detection task (OVD) has emerged, using image-text pairs for coaching and incorporating new class names at check time by associating them with the picture content material. By treating classes as textual content embeddings, open-vocabulary detectors can predict a variety of unseen objects. Varied methods corresponding to image-text pre-training, knowledge distillation, pseudo labeling, and frozen fashions, usually using convolutional neural network (CNN) backbones, have been proposed. With the rising reputation of vision transformers (ViTs), it is very important discover their potential for constructing proficient open-vocabulary detectors.

The present approaches assume the supply of pre-trained vision-language models (VLMs) and deal with fine-tuning or distillation from these fashions to deal with the disparity between image-level pre-training and object-level fine-tuning. Nevertheless, as VLMs are primarily designed for image-level duties like classification and retrieval, they don’t absolutely leverage the idea of objects or areas throughout the pre-training part. Thus, it could possibly be useful for open-vocabulary detection if we construct locality info into the image-text pre-training.

In “RO-ViT: Region-Aware Pretraining for Open-Vocabulary Object Detection with Vision Transformers”, offered at CVPR 2023, we introduce a easy technique to pre-train imaginative and prescient transformers in a region-aware method to enhance open-vocabulary detection. In imaginative and prescient transformers, positional embeddings are added to picture patches to encode details about the spatial place of every patch throughout the picture. Commonplace pre-training sometimes makes use of full-image positional embeddings, which doesn’t generalize properly to detection duties. Thus, we suggest a brand new positional embedding scheme, known as “cropped positional embedding”, that higher aligns with using area crops in detection fine-tuning. As well as, we substitute the softmax cross entropy loss with focal loss in contrastive image-text studying, permitting us to study from tougher and informative examples. Lastly, we leverage current advances in novel object proposals to boost open-vocabulary detection fine-tuning, which is motivated by the statement that current strategies usually miss novel objects throughout the proposal stage because of overfitting to foreground classes. We’re additionally releasing the code here.

Area-aware image-text pre-training

Present VLMs are skilled to match a picture as a complete to a textual content description. Nevertheless, we observe there’s a mismatch between the best way the positional embeddings are used within the current contrastive pre-training approaches and open-vocabulary detection. The positional embeddings are necessary to transformers as they supply the knowledge of the place every factor within the set comes from. This info is commonly helpful for downstream recognition and localization duties. Pre-training approaches sometimes apply full-image positional embeddings throughout coaching, and use the identical positional embeddings for downstream duties, e.g., zero-shot recognition. Nevertheless, the popularity happens at region-level for open-vocabulary detection fine-tuning, which requires the full-image positional embeddings to generalize to areas that they by no means see throughout the pre-training.

To deal with this, we suggest cropped positional embeddings (CPE). With CPE, we upsample positional embeddings from the picture dimension typical for pre-training, e.g., 224×224 pixels, to that typical for detection duties, e.g., 1024×1024 pixels. Then we randomly crop and resize a area, and use it because the image-level positional embeddings throughout pre-training. The place, scale, and side ratio of the crop is randomly sampled. Intuitively, this causes the mannequin to view a picture not as a full picture in itself, however as a area crop from some bigger unknown picture. This higher matches the downstream use case of detection the place recognition happens at region- relatively than image-level.

For the pre-training, we suggest cropped positional embedding (CPE) which randomly crops and resizes a area of positional embeddings as an alternative of utilizing the whole-image positional embedding (PE). As well as, we use focal loss as an alternative of the widespread softmax cross entropy loss for contrastive studying.

We additionally discover it useful to study from onerous examples with a focal loss. Focal loss permits finer management over how onerous examples are weighted than what the softmax cross entropy loss can present. We undertake the focal loss and substitute it with the softmax cross entropy loss in each image-to-text and text-to-image losses. Each CPE and focal loss introduce no further parameters and minimal computation prices.

Open-vocabulary detector fine-tuning

An open-vocabulary detector is skilled with the detection labels of ‘base’ classes, however must detect the union of ‘base’ and ‘novel’ (unlabeled) classes at check time. Regardless of the spine options pre-trained from the huge open-vocabulary knowledge, the added detector layers (neck and heads) are newly skilled with the downstream detection dataset. Present approaches usually miss novel/unlabeled objects within the object proposal stage as a result of the proposals are inclined to classify them as background. To treatment this, we leverage current advances in a novel object proposal technique and undertake the localization quality-based objectness (i.e., centerness rating) as an alternative of object-or-not binary classification rating, which is mixed with the detection rating. Throughout coaching, we compute the detection scores for every detected area because the cosine similarity between the area’s embedding (computed through RoI-Align operation) and the textual content embeddings of the bottom classes. At check time, we append the textual content embeddings of novel classes, and the detection rating is now computed with the union of the bottom and novel classes.

The pre-trained ViT spine is transferred to the downstream open-vocabulary detection by changing the worldwide common pooling with detector heads. The RoI-Align embeddings are matched with the cached class embeddings to acquire the VLM rating, which is mixed with the detection rating into the open-vocabulary detection rating.

Outcomes

We consider RO-ViT on the LVIS open-vocabulary detection benchmark. On the system-level, our greatest mannequin achieves 33.6 field average precision on uncommon classes (AP_r) and 32.1 mask AP_r, which outperforms the very best current ViT-based method OWL-ViT by 8.0 AP_r and the very best CNN-based method ViLD-Ens by 5.8 masks AP_r. It additionally exceeds the efficiency of many different approaches primarily based on information distillation, pre-training, or joint coaching with weak supervision.

RO-ViT outperforms each the state-of-the-art (SOTA) ViT-based and CNN-based strategies on LVIS open-vocabulary detection benchmark. We present masks AP on uncommon classes (APr) , aside from SOTA ViT-based (OwL-ViT) the place we present field AP.

Other than evaluating region-level illustration by open-vocabulary detection, we consider the image-level illustration of RO-ViT in image-text retrieval by the MS-COCO and Flickr30K benchmarks. Our mannequin with 303M ViT outperforms the state-of-the-art CoCa mannequin with 1B ViT on MS COCO, and is on par on Flickr30K. This reveals that our pre-training technique not solely improves the region-level illustration but additionally the worldwide image-level illustration for retrieval.

We present zero-shot image-text retrieval on MS COCO and Flickr30K benchmarks, and evaluate with dual-encoder strategies. We report recall@1 (top-1 recall) on image-to-text (I2T) and text-to-image (T2I) retrieval duties. RO-ViT outperforms the state-of-the-art CoCa with the identical spine.

RO-ViT open-vocabulary detection on LVIS. We solely present the novel classes for readability. RO-ViT detects many novel classes that it has by no means seen throughout detection coaching: “fishbowl”, “sombrero”, “persimmon”, “gargoyle”.

Visualization of positional embeddings

We visualize and evaluate the discovered positional embeddings of RO-ViT with the baseline. Every tile is the cosine similarity between positional embeddings of 1 patch and all different patches. For instance, the tile within the top-left nook (marked in purple) visualizes the similarity between the positional embedding of the situation (row=1, column=1) and people positional embeddings of all different places in 2D. The brightness of the patch signifies how shut the discovered positional embeddings of various places are. RO-ViT types extra distinct clusters at totally different patch places exhibiting symmetrical international patterns across the middle patch.

Every tile reveals the cosine similarity between the positional embedding of the patch (on the indicated row-column place) and the positional embeddings of all different patches. ViT-B/16 spine is used.

Conclusion

We current RO-ViT, a contrastive image-text pre-training framework to bridge the hole between image-level pre-training and open-vocabulary detection fine-tuning. Our strategies are easy, scalable, and simple to use to any contrastive backbones with minimal computation overhead and no enhance in parameters. RO-ViT achieves the state-of-the-art on LVIS open-vocabulary detection benchmark and on the image-text retrieval benchmarks, exhibiting the discovered illustration isn’t solely useful at region-level but additionally extremely efficient on the image-level. We hope this research will help the analysis on open-vocabulary detection from the attitude of image-text pre-training which might profit each region-level and image-level duties.

Acknowledgements

Dahun Kim, Anelia Angelova, and Weicheng Kuo performed this work and at the moment are at Google DeepMind. We wish to thank our colleagues at Google Analysis for his or her recommendation and useful discussions.