Cells are initially detected in frame t₀ and then tracked to frame t₁. Object queries are processed in the first layer of the decoder for object detection. In the subsequent decoder layers, track queries are concatenated with the object queries to perform tracking.

Similar Items

• Cell-TRACTR detects cells in the first frame at time t 0. Object queries predicted to be a “cell” are converted to track queries. For Track Group Denoising (TGD), random noise is added to the track queries to get noised track queries. For Query Denoising (QD), random noise is added to the ground truth bounding boxes to generate noised track queries. All queries are processed together by the decoder. Attention masks are used to prevent information leakage between the object and track queries, the noised track queries generated from TGD, and the noised track queries generated from QD. The gray boxes indicate information being blocked between sets of queries while colored boxes indicate the flow of information.
  by: Owen M. O’Connor (11965460)
  Published: (2025)
• Example illustrating how object and track queries interact using an attention map processed by Cell-TRACTR across two subsequent frames. The schematic displays the actual output alongside the output the from top matching object queries. Extracted from the self-attention mechanism in the decoder, the attention map highlights the strength of interactions among the queries. Lighter colors indicate stronger interactions. The track queries tend to focus on themselves, while the object queries attend to the track query that corresponds most closely in location.
  by: Owen M. O’Connor (11965460)
  Published: (2025)
• Visual representation of how Cell-TRACTR uses track queries to track cells and detect divisions.
  by: Owen M. O’Connor (11965460)
  Published: (2025)
• (A) Example illustrating the correlation between the final prediction of an object query and where the object query is derived from within the multi-scale features. Colored pixels represent multi-scale features that were classified as “cell” by the decoder. Black pixels represent multi-scale feature pixels that were classified as “no object” by the decoder. The colored pixels are overlaid onto the phase contrast image for visualization purposes. (B) A more complex example illustrating the correlation between the multi-scale features and object queries.
  by: Owen M. O’Connor (11965460)
  Published: (2025)
• The encoder employs deformable self-attention to transform multi-scale features into encoded multi-scale features. Query selection is used to derive the object queries from these encoded multi-scale features. Each encoded image feature undergoes classification, and the top-K features are selected as the content embeddings for the object queries. Additionally, bounding boxes are extracted from the top-K segmentation masks and are subsequently converted into positional embeddings. Following this, the object queries perform self-attention and deformable cross-attention with the encoded multi-scale features to produce the output embeddings. MLP, multi-layer perceptron.
  by: Owen M. O’Connor (11965460)
  Published: (2025)