Release 2.0.4 - Add img_size to inference, rename detector module

Author: orlandocloss

CHANGELOG.md

@@ -4,6 +4,17 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [2.0.4] - 2025-02-02
+
+### Added
+- **Configurable inference image size**: New `img_size` parameter in `inference()` to match training size (default: 60)
+
+### Changed
+- Renamed `insect_detector.py` to `detector.py` for cleaner module naming
+
+### Fixed
+- **Critical**: Inference now uses correct image size for classification. Previously hardcoded to 768x768 → 640, which caused poor accuracy when training with smaller sizes (e.g., 60px)
+
 ## [2.0.3] - 2025-01-28
 
 ### Added

README.md

@@ -153,6 +153,7 @@ results = bplusplus.inference(
     fps=None,               # None = process all frames
     backbone="resnet50",    # Must match training
     save_video=True,        # Set to False to skip video rendering (only CSV output)
+    img_size=60,            # Must match training
 )
 
 print(f"Detected {results['tracks']} tracks ({results['confirmed_tracks']} confirmed)")

notebooks/full_pipeline.ipynb

@@ -304,6 +304,7 @@
     "    fps=None,              # None = all frames\n",
     "    backbone=\"resnet50\",   # Must match training\n",
     "    save_video=True,       # Set to False to skip video rendering (only CSV output)\n",
+    "    img_size=60,           # Must match training\n",
     ")\n",
     "\n",
     "print(f\"Detected {results['tracks']} tracks ({results['confirmed_tracks']} confirmed)\")"

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "bplusplus"
-version = "2.0.3"
+version = "2.0.4"
 description = "A simple method to create AI models for biodiversity, with collect and prepare pipeline"
 authors = ["Titus Venverloo <tvenver@mit.edu>", "Deniz Aydemir <deniz@aydemir.us>", "Orlando Closs <orlandocloss@pm.me>", "Ase Hatveit <aase@mit.edu>"]
 license = "MIT"

src/bplusplus/insect_detector.py src/bplusplus/detector.pysrc/bplusplus/insect_detector.py renamed to src/bplusplus/detector.py

@@ -1,6 +1,6 @@
 """
-Insect Detection Backend Module
-===============================
+Detection Backend Module
+========================
 
 This module provides motion-based insect detection utilities used by the inference pipeline.
 It is NOT meant to be run directly - use inference.py instead.

src/bplusplus/inference.py

@@ -18,7 +18,7 @@
 import logging
 
 from .tracker import InsectTracker
-from .insect_detector import (
+from .detector import (
     DEFAULT_DETECTION_CONFIG,
     get_default_config,
     build_detection_params,
@@ -307,7 +307,7 @@ class VideoInferenceProcessor:
     and track-based prediction aggregation.
     """
 
-    def __init__(self, species_list, hierarchical_model_path, params, backbone="resnet50"):
+    def __init__(self, species_list, hierarchical_model_path, params, backbone="resnet50", img_size=60):
         """
         Initialize the processor.
         
@@ -316,7 +316,9 @@ def __init__(self, species_list, hierarchical_model_path, params, backbone="resn
             hierarchical_model_path: Path to trained model weights
             params: Detection parameters dict
             backbone: ResNet backbone ('resnet18', 'resnet50', 'resnet101')
+            img_size: Image size for classification (should match training)
         """
+        self.img_size = img_size
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         self.species_list = species_list
         self.params = params
@@ -354,8 +356,7 @@ def __init__(self, species_list, hierarchical_model_path, params, backbone="resn
         self.model.eval()
 
         self.transform = transforms.Compose([
-            transforms.Resize((768, 768)),
-            transforms.CenterCrop(640),
+            transforms.Resize((self.img_size, self.img_size)),
             transforms.ToTensor(),
             transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
         ])
@@ -1080,6 +1081,7 @@ def inference(
     backbone="resnet50",
     crops=False,
     save_video=True,
+    img_size=60,
 ):
     """
     Run inference on a video file.
@@ -1098,6 +1100,7 @@ def inference(
                   If model checkpoint contains backbone info, it will be used instead.
         crops: If True, save cropped frames for each classified track
         save_video: If True, save annotated and debug videos. Defaults to True.
+        img_size: Image size for classification (should match training). Default: 60.
     
     Returns:
         dict: Processing results with output file paths
@@ -1166,6 +1169,7 @@ def inference(
         hierarchical_model_path=hierarchical_model_path,
         params=params,
         backbone=backbone,
+        img_size=img_size,
     )
 
     try:
@@ -1238,6 +1242,8 @@ def main():
                        help='Save cropped frames for each classified track')
     parser.add_argument('--no-video', action='store_true',
                        help='Skip saving annotated and debug videos')
+    parser.add_argument('--img-size', type=int, default=60,
+                       help='Image size for classification (should match training, default: 60)')
 
     # Detection parameters (override config)
     defaults = DEFAULT_DETECTION_CONFIG
@@ -1315,6 +1321,7 @@ def main():
         backbone=args.backbone,
         crops=args.crops,
         save_video=not args.no_video,
+        img_size=args.img_size,
     )
 
     if result.get("success"):