The skill to detect objects within the visible world is essential for pc vision and machine intelligence, enabling purposes like adaptive autonomous brokers and versatile buying programs. However, fashionable object detectors are restricted by the handbook annotations of their coaching information, leading to a vocabulary dimension considerably smaller than the huge array of objects encountered in actuality. To overcome this, the open-vocabulary detection activity (OVD) has emerged, using image-text pairs for coaching and incorporating new class names at take a look at 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. Various strategies comparable to image-text pre-training, information distillation, pseudo labeling, and frozen fashions, typically using convolutional neural community (CNN) backbones, have been proposed. With the rising recognition of vision transformers (ViTs), it is very important discover their potential for constructing proficient open-vocabulary detectors.
The current approaches assume the provision of pre-trained vision-language fashions (VLMs) and give attention to fine-tuning or distillation from these fashions to handle the disparity between image-level pre-training and object-level fine-tuning. However, as VLMs are primarily designed for image-level duties like classification and retrieval, they don’t totally leverage the idea of objects or areas through the pre-training part. Thus, it may very well be helpful for open-vocabulary detection if we construct locality data 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 methodology to pre-train vision transformers in a region-aware method to enhance open-vocabulary detection. In vision transformers, positional embeddings are added to picture patches to encode details about the spatial place of every patch inside the picture. Standard 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, referred to as “cropped positional embedding”, that higher aligns with the usage of area crops in detection fine-tuning. In addition, we substitute the softmax cross entropy loss with focal loss in contrastive image-text studying, permitting us to be taught from tougher and informative examples. Finally, we leverage latest advances in novel object proposals to reinforce open-vocabulary detection fine-tuning, which is motivated by the statement that current strategies typically miss novel objects through the proposal stage because of overfitting to foreground classes. We are additionally releasing the code right here.
Region-aware image-text pre-training
Existing VLMs are educated to match a picture as a complete to a textual content description. However, 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 ingredient within the set comes from. This data is usually 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. However, 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 through the pre-training.
To handle 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- moderately than image-level.
 |
| For the pre-training, we suggest cropped positional embedding (CPE) which randomly crops and resizes a area of positional embeddings as a substitute of utilizing the whole-image positional embedding (PE). In addition, we use focal loss as a substitute of the widespread softmax cross entropy loss for contrastive studying. |
We additionally discover it helpful to be taught 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. Both CPE and focal loss introduce no additional parameters and minimal computation prices.
Open-vocabulary detector fine-tuning
An open-vocabulary detector is educated with the detection labels of ‘base’ classes, however must detect the union of ‘base’ and ‘novel’ (unlabeled) classes at take a look at time. Despite the spine options pre-trained from the huge open-vocabulary information, the added detector layers (neck and heads) are newly educated with the downstream detection dataset. Existing approaches typically 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 latest advances in a novel object proposal methodology and undertake the localization quality-based objectness (i.e., centerness rating) as a substitute of object-or-not binary classification rating, which is mixed with the detection rating. During coaching, we compute the detection scores for every detected area because the cosine similarity between the area’s embedding (computed by way of RoI-Align operation) and the textual content embeddings of the bottom classes. At take a look at 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. |
Results
We consider RO-ViT on the LVIS open-vocabulary detection benchmark. At the system-level, our greatest mannequin achieves 33.6 field common precision on uncommon classes (APr) and 32.1 masks APr, which outperforms the very best current ViT-based method OWL-ViT by 8.0 APr and the very best CNN-based method ViLD-Ens by 5.8 masks APr. It additionally exceeds the efficiency of many different approaches based mostly 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) , besides for SOTA ViT-based (OwL-ViT) the place we present field AP. |
Apart from 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 methodology not solely improves the region-level illustration but in addition the worldwide image-level illustration for retrieval.
 |
| We present zero-shot image-text retrieval on MS COCO and Flickr30K benchmarks, and examine 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 examine the realized positional embeddings of RO-ViT with the baseline. Each 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 crimson) visualizes the similarity between the positional embedding of the placement (row=1, column=1) and people positional embeddings of all different areas in 2D. The brightness of the patch signifies how shut the realized positional embeddings of various areas are. RO-ViT types extra distinct clusters at completely different patch areas exhibiting symmetrical world patterns across the heart patch.
 |
| Each 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 improve 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 realized illustration is just not solely helpful at region-level but in addition extremely efficient on the image-level. We hope this examine might 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 are actually at Google DeepMind. We wish to thank our colleagues at Google Research for their recommendation and useful discussions.