kelseylmartin/GFIT_v3

Model Card: GFIT v3 Fish Tracking Pipeline

Model Details

• Model Name: GFIT v3 Tracker
• Model Type: Fused Object Detection and Kinematic Tracking Pipeline
• Release Date: August 2026[cite: 1]
• Core Architecture: The pipeline fuses a high-resolution visual detector (sefsc_rgb_1728_groups) with a motion detector[cite: 1]. It utilizes the ByteTrack algorithm for kinematic tracking instead of appearance-based Re-ID[cite: 1].

Uses

• Primary Intended Use: Identifying and tracking dense schools of fish in underwater stationary benthic camera arrays.
• Specific Capabilities: The tracker is heavily optimized to maintain fish IDs during occlusions (such as fish swimming behind structural camera bars). It implements "track averaging" to stabilize species classification across a track's lifespan, ignoring unknown classifications[cite: 1].

Evaluation Data

• Dataset: SEFSC test set[cite: 1].
• Size: Evaluated across 121 total sequences, with a 13-sequence subset specifically used for group and species classification[cite: 1].
• Data Split Limitations: While exact overlap between training and testing is zero, the dataset uses a clip-level split rather than a deployment-level split[cite: 1]. 50 of the 97 test clips are sibling segments of recordings whose other segments appear in the training data[cite: 1].

Evaluation Results

Detection & Classification

• Detector Fusion: Fusing the sefsc_rgb_1728_groups and motion detectors beat every single standalone detector, achieving an AP@50 of 0.6659 compared to the single-best score of 0.6380[cite: 1].
• Track Averaging: Implementing track averaging is the single strongest lever for performance, increasing group mAP@50 by +0.038 over the best detection-level method (reaching 0.474)[cite: 1].
• Averaging Configuration: Track averaging is shipped using required_states = 3, with the average weighted and scaled by confidence[cite: 1].
• Species Identification: The GFIT v3 pipeline uses a blended-group prior of 0.65, achieving a species mAP@50 of 0.320, which is a marked improvement over the older v2.5 baseline of 0.241[cite: 1].

Tracking Performance (ByteTrack)

• Tracker Dominance: Motion-only kinematic trackers entirely dominated appearance-based Re-ID models (like BoT-SORT and DeepSORT) on this dataset[cite: 1].
• The Operating Point: The shipped hi.30 ByteTrack configuration balances quality and coverage[cite: 1]. Evaluated at an output confidence of 0.42, it achieves a HOTA score of 0.5215[cite: 1].
• Track Coverage: This configuration captures a track_pd of 0.786, meaning 78.6% of all ground-truth tracks are successfully detected by the model[cite: 1].

Pipeline Configuration Settings

The following hyperparameter settings dictate the ByteTrack pipeline logic, extracted directly from the tuned GFIT v3 configuration file.

• Tracker Type: bytetrack
• Standard Weight Position (std_weight_position): 0.050630267994044076
• Standard Weight Velocity (std_weight_velocity): 0.1
• Track High Threshold (high_thresh): 0.30
• Track Low Threshold (low_thresh): 0.05
• New Track Threshold (new_track_thresh): 0.30
• Track Buffer (track_buffer): 10
• Match Threshold (match_thresh): 0.90
• Second Match Threshold (second_match_thresh): 0.90
• Unconfirmed Match Threshold (unconfirmed_match_thresh): 0.90