# GC Rebuttal - ECCV 2020

# R1

Questions:

1. Experimenting with more ways of generating the global context (besides dot-product).
2. More sophisticated way of aggregating context across frames (besides averaging). Reviewer suspects it might be causing low-performance on long clips on YouTube-VOS (I don't think he/she's correct about this).
3. Key and value are bad names.
4. High score on YouTube-VOS

# R2

Questions:

1. Comparison vs other memory mechanisms, e.g. RNNs, LSTMs.
2. Acknowledging early, non-deep works.
3. GC with soft aggregation or STM without soft aggregation for fair comparision.

# R3

Questions:

1. Limited novelty. Only differnce is using GC module to compress the memory features.
2. Making the memory adaptive rather than simply an average.
3. Ablation studies on the efficacy of weighted summation. E.g. avg_pool vs. max_pool.