Add phases

README.md

@@ -1,3 +1,50 @@
 # Pose Estimation Task Force
 
 Placeholder repository for issues related to the pose estimation task force.
+
+
+
+## Phase Plan
+
+The mission of this project is quite extensive in scope.
+As such, we are carefully planning each stage as limited, achievable steps, which will gradually accumulate the capabilities of the system.
+We thus state the following milestones with expected deliverable dates and subtasks.
+
+Dataset-specific RAG-style methods will be assessed and used throughout each phase.
+Individual models will be trained per video and per dataset of consistent videos.
+Mixtures of experts will be made across datasets.
+Long-running cumulative models will be trained continuously over the entire collection.
+
+### Phase 1 (December 2026)
+
+- Collect a centralized collection of videos with top-down or bottom-up fixed camera field of views of caged rodents.
+- Transcode all videos in the collection to a common physical or coordinate space.
+- Work with the scientific community to define rigorous, well-documented, and anatomically justified standard skeleton definitions (at both coarse- and fine- levels of detail).
+- Label 1,000,000 frames from the standard transcoding of the data collection using the standardized collection.
+- Offer automatically generated ethogram methods from the inferred pose series for each model.
+  - Ideally, also compare to some manually annotated ethograms for ground truth benchmark comparison.
+
+### Phase 2 (June 2027)
+
+- Expand data collection and foundation model training and inference to synchronized multi-camera cases while maintaining previous restrictions on the environment.
+- Reach 2,000,000 labeled frames across all collections.
+- Analyze and prioritize frame labeling based on predicted information value (not necessarily uniform over time).
+
+### Phase 3 (December 2027)
+
+- Expand data collection and foundation model training and inference to multi-subject single-camera (top-down/bottom-up) cases while maintaining previous restrictions on field of view and environment.
+- Reach 3,000,000 labeled frames across all collections.
+
+### Phase 4 (June 2028)
+
+- Expand data collection and foundation model training and inference to multi-subject, multi-camera cases while maintaining previous restrictions on the environment.
+- Reach 4,000,000 labeled frames across all collections.
+
+### Phase 5 (December 2028)
+
+- Expand data collection and foundation model training and inference to non-caged (open) environments.
+  - By its nature, this will include a variety of fields of view.
+  - Start with single subject cases where possible and expand later to multi-subject.
+- Reach 5,000,000 labeled frames across all collections.
+
+The hope is that phases 1-4 will make the extension to phase 5 relatively smooth.