The image showcases a scientific research poster presented at CVPR (Conference on Computer Vision and Pattern Recognition), scheduled for June 17-21, 2024, in Seattle, WA. The title of the work involves understanding and developing open vocabularies, credited to C. Cottereau and Wei Tsang Ooi. The poster elaborates on various methodologies, experiments, and analyses related to OpenESS (Event Segmentation System).

Key sections include:
- Comparative & Ablation Study, highlighting that OpenESS has achieved state-of-the-art (SoTA) results under various setups: zero-shot, fully-supervised, and open-vocabulary.
- Detailed results are presented in tabular form under "Event Representation Learning" and comparative ESS settings.
- Graphs and charts illustrate the performance metrics under different conditions, emphasizing the improvement in results through OpenESS.
- The importance of factors such as distillation strength, linear scaling, and label efficiency are discussed.

The poster also indicates the implications for future ESS system development, emphasizing scalability, robustness, and accuracy to ensure safety. The presence of attendees interacting with the content highlights the engaging nature of the presentation.
Text transcribed from the image:
standing with Open Vocabularies
R. Cottereau        Wei Tsang Ooi

Comparative & Ablation Study
OpenESS achieved SoTA results under zero-shot, fully-supervised, and open-vocabulary ESS setups.

Tab. Event Representation Learning
Method                Venue          Backbone      OV    DDD17    DSEC
Howo3D (Li)             ICCV'23      VIT-B/16        X           X                 X
BEV (Huang)            ICCV'23       Swin-B            X           X                 X
PPT (Liu)                    ICCV'23        VIT-B/16        X           X                 X
ERMF (Kang)          RAL'23          VIT-B/16        X           X                 X
PIDNet (Xu)              TITS'23         ResNet-50    X           X                 X
RVSA (Ling)               ICCV'23         ResNet-50    X           X                 X
ESSFormer (Zhao)  WACV'23       ResNet-50    X           X                 X
OpenESS                 Ours               EViT-B        67.01      60.37       51.05

OpenESS exhibits better results than other event representation learning methods in literature.
We unveil important factors of adapting better image and text knowledge to the event network, strength of distillation, linear scaling, and label efficiency.

Impact on distillation strength

Fig. Cross-Domain ESS Pretraining
OpenESS shed lights on future development of more scalable ESS systems in the real world.
By incorporating the image-text knowledge, we anticipate event perception models to be robust and accurate to ensure safety.

CVPR
JUNE 17-21, 2024
SEATTLE, WA