README.md

Fast R-CNN

This implementation trains its backbone from scratch. Detector images are resized to 512x512 and the stride-16 backbone produces a 32x32 feature map. The default global batch is 4; set DETECTOR_BATCH_SIZE=N only after measuring a real training step, and keep it divisible by the number of visible replicas.

COCO 2017, Pascal VOC 2012, and ImageNet CLS-LOC have separate class heads, proposal caches, checkpoints, logs, and inference outputs. VOC uses its labelled train.txt and val.txt splits.

Required selective search

Fast R-CNN requires the OpenCV contrib build. The program stops if selective search is missing instead of silently training with low-recall grid proposals.

/home/shuvrajeet/miniconda3/envs/research/bin/python -m pip uninstall opencv-python
/home/shuvrajeet/miniconda3/envs/research/bin/python -m pip install opencv-contrib-python==4.13.0.92

Measure proposal recall before building a complete cache. On the local validation check, fast mode reached 0.77 COCO recall@0.5 and finished the complete cache at 3.2 images/second; quality mode reached 0.83 but was much slower. Fast mode and 4,000 proposal slots are therefore the practical defaults. Set FAST_RCNN_PROPOSAL_MODE=quality only when the added recall is worth rebuilding the cache.

cd /home/shuvrajeet/Documents/TensorflowAI/ComputerVision/ObjectDetection/2_FastRCNN
PYTHON=/home/shuvrajeet/miniconda3/envs/research/bin/python
$PYTHON train.py --detector coco --proposal-report
$PYTHON train.py --detector voc --proposal-report
FAST_RCNN_PROPOSAL_MODE=quality $PYTHON train.py --detector coco --proposal-report

The proposal mode, OpenCV version, input size, and sampling settings are part of the cache fingerprint. Old caches remain untouched and are not reused after these changes.

Training and evaluation

Training uses 64 ROIs per image, at most 25% foreground, hard background proposals from the 0.1-0.5 IoU interval, and exact ground-truth boxes only in the training sampler. Validation and inference use genuine selective-search proposals. Horizontal flips transform the image, ROIs, and regression targets together.

Run independent COCO and VOC training jobs sequentially with one command:

sh run.sh --reset --epochs 100

Arguments are forwarded to both runs. If the COCO run fails, the VOC run is not started.

$PYTHON train.py --detector coco --prepare-cache
$PYTHON train.py --detector coco --reset --epochs 100
$PYTHON train.py --detector coco --continue --epochs 100
$PYTHON train.py --detector coco --evaluate

$PYTHON train.py --detector voc --prepare-cache
$PYTHON train.py --detector voc --reset --epochs 100
$PYTHON train.py --detector voc --evaluate

Use --reset for the first run with the 512x512 stride-16 architecture. Detector training evaluates a fixed 200-image validation subset every epoch and keeps one best checkpoint: COCO uses subset mAP and VOC/ImageNet use AP50. Inference scores each foreground class independently, applies class-wise NMS, and returns at most 100 detections per image.

Persistent proposal caches live under cache/FastRCNN/<dataset>/<split>/<fingerprint>/. Use --rebuild-cache only to replace the currently selected generation, --no-cache for lazy proposal generation, and --cache-workers 1 for serial preparation. --full-reports adds TensorBoard and model diagrams; validation metrics and 4x4 inspection grids are always produced for detectors.