Skip to content
TheChatArchiveThe Chat Archive
Googlepalm-2· GoogleHugging Face Arena Human Preference 55K

"CoTracker: It is Better to Track Together": Proposes CoTracker, a new method for…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with vicuna-7b’s answer.
  1. User

    "CoTracker: It is Better to Track Together":

    Proposes CoTracker, a new method for jointly tracking multiple points in videos over long time periods. Uses a transformer architecture to track groups of points simultaneously and leverage their correlation. Has interleaved time and group attention blocks to make this computationally feasible. Can track arbitrary points selected at any spatial location and time in the video. Initializes tracks trivially by assuming points stay stationary. Operates in a sliding window fashion to handle long videos. Unrolled training loop allows learning to propagate information between windows. Evaluated on multiple benchmarks including TAP-Vid, BADJA, and FastCapture. Outperforms prior state-of-the-art in point tracking, especially for long tracks. Key benefits are accounting for correlation between points, tracking jointly, and unrolled window training. Allows tracking through longer occlusions. Limitations include quadratic complexity in number of points and inability to handle occlusions longer than the window size. In summary, the main contributions are proposing a new transformer-based architecture for joint point tracking that outperforms previous approaches, especially on long videos and tracks. The core ideas are modeling inter-point correlations, iterative updating of tracks, and unrolled window training.

  2. Google Assistant

    Great summary! I'm glad you were able to identify the key points of the paper.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.

Identical first prompt