multispectral_uav_detection: weights

Checkpoints for marymikkey/multispectral_uav_detection (YOLO11s, RGB / IR / RGBT mid-fusion P3, built on the YOLOv11-RGBT fork, commit 9cc2e208).

The folder layout matches DIPLOMA_ROOT in the code repo, so the evaluation and transfer scripts find the weights without path changes:

hf download marymikkey/multispectral_uav_detection --local-dir $DIPLOMA_ROOT

Path	Model
pretrain/runs/yolo11s_rgb_pretrain	RGB pretrain (unpaired data), grayscale input
pretrain/runs/yolo11s_gray_ir_pretrain_100ep	IR pretrain (unpaired data)
train/runs/antiuav_resync/yolo11s_rgb_from_scratch	RGB-only, Anti-UAV-RGBT, grayscale input
train/runs/antiuav_resync/yolo11s_ir_from_scratch	IR-only, Anti-UAV-RGBT
train/runs/antiuav_resync/yolo11s_rgbt_midfusion_p3_from_scratch	RGBT mid-fusion from scratch
train/runs/antiuav_resync/yolo11s_rgbt_midfusion_p3_rgb_ir_pretrained_{rgb,ir}_neck_head_freeze_backbones	RGBT, transferred frozen backbones, RGB or IR neck/head
train/runs/antiuav_resync/*_poly_finetune_crop_x287_1248	the five models above fine-tuned on field data (1248x1248)

Input modes and colour retraining

The RGB pretrain and the RGB-only Anti-UAV model above were trained without setting the fork's use_simotm argument, so the default mode was used: the visible frame is read as grayscale and the three channels are filled with the gray image and two blurred copies. The RGBT models read the visible frame in colour.

Both models were retrained with colour input (use_simotm="RGB"), nothing else changed, and both RGBT transfer variants were re-initialised from the colour RGB pretrain:

Path	Model
pretrain/runs/yolo11s_rgb_pretrain_color	RGB pretrain, colour input
train/runs/antiuav_resync/yolo11s_rgb_from_scratch_color	RGB-only, colour input
train/runs/antiuav_resync/yolo11s_rgbt_midfusion_p3_rgb_color_ir_pretrained_{rgb,ir}_neck_head_freeze_backbones	RGBT transfer from the colour RGB pretrain
train/runs/antiuav_resync/yolo11s_rgbt_midfusion_p3_rgb_color_{rgb,ir}_neck_head_freeze_backbones_poly_finetune_crop_x287_1248	the two colour transfer models fine-tuned on field data

Results and comparison: see the code repo README. Each best.pt stores its training arguments. SHA256SUMS and SHA256SUMS_colour list checksums.

LLVIP

RGB-only, IR-only and RGBT mid-fusion P3 trained on LLVIP (person, night street scenes) with the Anti-UAV settings; the RGB model uses colour input.

Path	Model
train/runs/llvip/yolo11s_{rgb,ir,rgbt}_llvip	RGB-only / IR-only / RGBT on LLVIP

Results: experiments/llvip/README.md in the code repo. SHA256SUMS_llvip lists checksums.