tidalove/cfc-track-instruction

CFC Track & Track-Correction Instruction Data

Fish tracking and multi-turn track-correction annotations on the Caltech Fish Counting sonar videos, in the Molmo2 video-track-instruction format. Companion to the Molmo2 codebase's olmo/data/cfc_hf_datasets.py loader classes.

Videos are 6 fps clips; annotations are stored at 2 fps: only native frames with frame % 3 == 0 are annotated, frame keeps the native 6 fps index, and time = frame / 6. Inline masks follow the same convention — mask entry i corresponds to native frame 3 * i.

Configs

config	format	splits
cfc_track	track (one row per clip, "track all fish")	train / validation
cfc_target	track (one row per query; qid 0 = all fish, others = referred subsets)	train / validation
cfc_synthetic_correction_full / _vague / _wrong_only / _no_info	correction, synthetically corrupted step 0	train / validation
cfc_synthetic_correction_incomplete	correction, partial-fix target (GT = final step, NOT the full video GT)	train / validation
cfc_correction_real_full_easy / _vague_easy / _no_info_easy	correction, real model-prediction step 0 (easy set)	train / validation
cfc_correction_real_wrong_only_easy	correction, real (easy set)	train only
cfc_correction_real_full_hard / _vague_hard / _wrong_only_hard / _no_info_hard	correction, real model-prediction step 0 (hard set)	train / validation
cfc_correction_real_yolo_full / _vague / _wrong_only / _no_info	correction, YOLO-SORT tracker step 0	validation only
cfc_text	text-only correction (no masks, no video input at train time)	train / validation

Schema

Track configs: id, video, clip, video_dataset, task, expression, qid, prepend, width, height, fps (6), sampling_fps (2), n_frames, start_frame, end_frame, mask_id, obj_id, anno_id, frame_trajectories, masks.

frame_trajectories: list of {frame, time, points: [{id, point: [x, y], occluded}]}. Point id indexes into mask_id/obj_id slots.

Correction configs replace prepend/anno_id/frame_trajectories with confidence and turns: list of {correction_step, prompt, frame_trajectories} — a multi-turn conversation where step 0 is the (possibly corrupted or model-predicted) starting tracks and later steps are correction targets. Prompts are stored raw with native frame references; loaders rewrite them to timestamps (frame N -> {N/fps}s). Per-turn point ids are per-step slot indices (sorted track ids of that step).

masks: list of {object_id, masks: [RLE | null]} — pycocotools RLE ({size: [h, w], counts}), null = object absent. For most correction configs these are the base-video GT masks (slot order = sorted video track ids, which may differ from per-turn point slots — consumers should match by mask content, as the Molmo2 eval does). For cfc_synthetic_correction_incomplete they encode the final correction step per trajectory.

Usage

import datasets
ds = datasets.load_dataset("tidalove/cfc-track-instruction", "cfc_track", split="validation")

With the Molmo2 codebase, olmo/data/cfc_hf_datasets.py downloads all configs, rehydrates local MasksRLE/ files, and (given frames) encodes videos: place frames at $MOLMO_DATA_DIR/video_datasets/video_track/CFC/JPEGImages/{video_id}/*.jpg, then call download() on any of the loader classes (or use the registered cfc_hf_* dataset names).

Built by scripts/build_cfc_hf_dataset.py; maintenance guide in docs/cfc_hf_dataset.md.