Exploration on REID and trajectory embeddings

[hector6298]

Approach 1

• Search for surveys on REID

Examples:

• https://ieeexplore.ieee.org/abstract/document/9336268
• https://arxiv.org/abs/2303.11332

Approach 1A (One embedding per trajectory + average)

• Embeddings with different REID models
  • Explore different architectures

Approach 1B (One embedding per trajectory + Other approach?)

• Search trajectory generation
  • Current approach is just the mean vector. Search for other methodologies

Examples

• MLP, CNN + Global AvgPool for trajectory embedding generation
• Concatenation of single detection embeddings

Approach 2 (Multiple embeddings per trajectory)

• DeepSort appearance matching with galleries (search for paper, both nodes and edges)

• Have all of the embeddings, not just the average. Inspect the optimal way of producing edge embeddings (Maybe the least distance)

• What do we want?

• Options available

  • Create a document with options for implementation and experimentation
  • List and describe papers

Approach 3 (Let the GNN itself improve the discriminability of the embeddings)

• based on baseline REID works that uses cross entropy + triplet loss
  • We may be able to use triplet loss + cross entropy on link prediction (possibly add node prediction as well) in a multi-task learning style.
  • This approach can also use approach 3 as an overall enhancement of the baseline methodology

[hector6298]

• Have a document with the outline of the experiments to launch experiments as we work on other approaches