Spatio-temporal graph for video captioning with knowledge distillation
A method for scene perception using video captioning based on a spatio-temporal graph model is described. The method includes decomposing the spatio-temporal graph model of a scene in input video into a spatial graph and a temporal graph. The method also includes modeling a two branch framework having an object branch and a scene branch according to the spatial graph and the temporal graph to learn object interactions between the object branch and the scene branch. The method further includes transferring the learned object interactions from the object branch to the scene branch as privileged information. The method also includes captioning the scene by aligning language logits from the object branch and the scene branch according to the learned object interactions.
1. A method for scene perception using video captioning based on a spatio-temporal graph model, comprising:
decomposing the spatio-temporal graph model of a scene in input video into a spatial graph and a temporal graph;
modeling a two branch framework having an object branch and a scene branch according to the spatial graph and the temporal graph to learn object interactions between the object branch and the scene branch;
transferring the learned object interactions from the object branch to the scene branch as privileged information; and
captioning the scene by aligning language logits from the object branch and the scene branch according to the learned object interactions.
2. The method of claim 1 , in which captioning comprises decoding the aligned language logits between the object branch and the scene branch to form a caption for the scene.
3. The method of claim 2 , in which language decoding is performed through a transformer network to obtain a final text description as the caption for the scene.
4. The method of claim 1 , in which transferring further comprises encoding a relationship between objects in the object branch and the scene in the scene branch.
5. The method of claim 1 , further comprising planning a trajectory of an ego vehicle according to perception of the scene from the video captured by the ego vehicle.
6. The method of claim 1 , further comprising:
modeling the object branch in a first neural network; and
modeling the scene branch in a second neural network.
7. The method of claim 1 , further comprising classifying the input video according to the captioning of the scene.
8. The method of claim 1 , in which the spatial graph is used to capture interactions among spatially related objects of the video.
9. The method of claim 1 , further comprising incorporating video information in the spatial graph by connecting objects using their normalized intersection over union (IoU) value for measuring a spatial connectivity between the objects.
10. The method of claim 1 , in which the object branch captures space-time object interaction information via the spatio-temporal graph model, and the scene branch provides additional global context missing from local object features.
11. A non-transitory computer-readable medium having program code recorded thereon for scene perception using video captioning based on a spatio-temporal graph model, the program code being executed by a processor and comprising:
program code to decompose the spatio-temporal graph model of a scene in input video into a spatial graph and a temporal graph;
program code to model a two branch framework having an object branch and a scene branch according to the spatial graph and the temporal graph to learn object interactions between the object branch and the scene branch;
program code to transfer the learned object interactions from the object branch to the scene branch as privileged information; and
program code to caption the scene by aligning language logits from the object branch and the scene branch according to the learned object interactions.
12. The non-transitory computer-readable medium of claim 11 , in which the program code to caption comprises program code to decode the aligned language logits between the object branch and the scene branch to form a caption for the scene, in which language decoding is performed through a transformer network to obtain a final text description as the caption for the scene.
13. The non-transitory computer-readable medium of claim 11 , in which the program code to transfer further comprises program code to encode a relationship between objects in the object branch and the scene in the scene branch.
14. The non-transitory computer-readable medium of claim 11 , further comprising program code to plan a trajectory of an ego vehicle according to perception of the scene from video captions of the video captured by the ego vehicle.
15. The non-transitory computer-readable medium of claim 11 , further comprising:
program code to model the object branch in a first neural network; and
program code to model the scene branch in a second neural network.
16. The non-transitory computer-readable medium of claim 11 , further comprising program code to classify the input video according to video captions of the scene.
17. The non-transitory computer-readable medium of claim 11 , further comprising program code to incorporate video information in the spatial graph by connecting objects using their normalized intersection over union (IoU) value for measuring a spatial connectivity between the objects, in which the spatial graph is used to capture interactions among spatially related objects of the video.
18. The non-transitory computer-readable medium of claim 11 , in which the object branch captures space-time object interaction information via the spatio-temporal graph model, and the scene branch provides additional global context missing from local object features.
19. A system for scene perception using video captioning based on a spatio-temporal graph model, the system comprising:
a spatio temporal graph module configured to program code to decompose the spatio-temporal graph model of a scene in input video into a spatial graph and a temporal graph;
a knowledge distillation module configured to model a two branch framework having an object branch and a scene branch according to the spatial graph and the temporal graph to learn object interactions between the object branch and the scene branch and transfer the learned object interactions from the object branch to the scene branch as privileged information; and
a video captioning module configured to caption the scene by aligning language logits from the object branch and the scene branch according to the learned object interactions.
20. The system of claim 19 , further comprising:
a scene comprehension module configured to comprehend the scene from the video captured by an ego vehicle according to video captioning of the scene; and
a planner module configured to plan a trajectory of the ego vehicle according to a comprehension of the scene from the video captured by the ego vehicle based on the scene comprehension module.